v1.15.1

- Update smartai to v0.13.2 (native tool calling support)
- Update smartagent to v1.5.1 (useNativeToolCalling option)
- Enable think: true for GPT-OSS reasoning mode in Ollama config
- Enable useNativeToolCalling: true in DualAgentOrchestrator
- Simplify driver system message (native tools don't need XML instructions)

Native tool calling uses Ollama's built-in Harmony format parser
instead of requiring XML generation, which is more efficient for GPT-OSS models.

Dockerfile_nanonets_vllm_gpu_VRAM10GB

@@ -0,0 +1,34 @@
+# Nanonets-OCR2-3B Vision Language Model
+# Based on Qwen2.5-VL-3B, fine-tuned for document OCR (Oct 2025 release)
+# Improvements over OCR-s: better semantic tagging, LaTeX equations, flowcharts
+# ~12-16GB VRAM with 30K context, outputs structured markdown with semantic tags
+#
+# Build: docker build -f Dockerfile_nanonets_vllm_gpu_VRAM10GB -t nanonets-ocr .
+# Run:   docker run --gpus all -p 8000:8000 -v ht-huggingface-cache:/root/.cache/huggingface nanonets-ocr
+
+FROM vllm/vllm-openai:latest
+
+LABEL maintainer="Task Venture Capital GmbH <hello@task.vc>"
+LABEL description="Nanonets-OCR2-3B - Document OCR optimized Vision Language Model"
+LABEL org.opencontainers.image.source="https://code.foss.global/host.today/ht-docker-ai"
+
+# Environment configuration
+ENV MODEL_NAME="nanonets/Nanonets-OCR2-3B"
+ENV HOST="0.0.0.0"
+ENV PORT="8000"
+ENV MAX_MODEL_LEN="30000"
+ENV GPU_MEMORY_UTILIZATION="0.9"
+
+# Expose OpenAI-compatible API port
+EXPOSE 8000
+
+# Health check - vLLM exposes /health endpoint
+HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=5 \
+    CMD curl -f http://localhost:8000/health || exit 1
+
+# Start vLLM server with Nanonets-OCR2-3B model
+CMD ["--model", "nanonets/Nanonets-OCR2-3B", \
+     "--trust-remote-code", \
+     "--max-model-len", "30000", \
+     "--host", "0.0.0.0", \
+     "--port", "8000"]

build-images.sh

@@ -13,46 +13,38 @@ NC='\033[0m' # No Color
 
 echo -e "${BLUE}Building ht-docker-ai images...${NC}"
 
-# Build GPU variant
+# Build MiniCPM-V 4.5 GPU variant
 echo -e "${GREEN}Building MiniCPM-V 4.5 GPU variant...${NC}"
 docker build \
-    -f Dockerfile_minicpm45v_gpu \
+    -f Dockerfile_minicpm45v_ollama_gpu_VRAM9GB \
     -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v \
     -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v-gpu \
     -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:latest \
     .
 
-# Build CPU variant
-echo -e "${GREEN}Building MiniCPM-V 4.5 CPU variant...${NC}"
+# Build Qwen3-VL GPU variant
+echo -e "${GREEN}Building Qwen3-VL-30B-A3B GPU variant...${NC}"
 docker build \
-    -f Dockerfile_minicpm45v_cpu \
-    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v-cpu \
+    -f Dockerfile_qwen3vl_ollama_gpu_VRAM20GB \
+    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:qwen3vl \
     .
 
-# Build PaddleOCR-VL GPU variant
-echo -e "${GREEN}Building PaddleOCR-VL GPU variant...${NC}"
+# Build Nanonets-OCR GPU variant
+echo -e "${GREEN}Building Nanonets-OCR-s GPU variant...${NC}"
 docker build \
-    -f Dockerfile_paddleocr_vl_gpu \
-    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl \
-    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl-gpu \
-    .
-
-# Build PaddleOCR-VL CPU variant
-echo -e "${GREEN}Building PaddleOCR-VL CPU variant...${NC}"
-docker build \
-    -f Dockerfile_paddleocr_vl_cpu \
-    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl-cpu \
+    -f Dockerfile_nanonets_vllm_gpu_VRAM10GB \
+    -t ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:nanonets-ocr \
     .
 
 echo -e "${GREEN}All images built successfully!${NC}"
 echo ""
 echo "Available images:"
-echo "  MiniCPM-V 4.5:"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v (GPU)"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v-cpu (CPU)"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:latest (GPU)"
+echo "  MiniCPM-V 4.5 (Ollama, ~9GB VRAM):"
+echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:minicpm45v"
+echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:latest"
 echo ""
-echo "  PaddleOCR-VL (Vision-Language Model):"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl (GPU/vLLM)"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl-gpu (GPU/vLLM)"
-echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:paddleocr-vl-cpu (CPU)"
+echo "  Qwen3-VL-30B-A3B (Ollama, ~20GB VRAM):"
+echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:qwen3vl"
+echo ""
+echo "  Nanonets-OCR-s (vLLM, ~10GB VRAM):"
+echo "  - ${REGISTRY}/${NAMESPACE}/${IMAGE_NAME}:nanonets-ocr"

changelog.md

@@ -1,5 +1,73 @@
 # Changelog
 
+## 2026-01-20 - 1.15.1 - fix(tests)
+enable progress events in invoice tests and bump @push.rocks/smartagent devDependency to ^1.5.4
+
+- Added an onProgress handler in test/test.invoices.nanonets.ts to log progress events (console.log(event.logMessage)) so tool calls and progress are visible during tests.
+- Bumped devDependency @push.rocks/smartagent from ^1.5.2 to ^1.5.4 in package.json.
+
+## 2026-01-20 - 1.15.0 - feat(tests)
+integrate SmartAi/DualAgentOrchestrator into extraction tests and add JSON self-validation
+
+- Integrate SmartAi and DualAgentOrchestrator into bankstatement and invoice tests to perform structured extraction with streaming
+- Register and use JsonValidatorTool to validate outputs (json.validate) and enforce validation before task completion
+- Add tryExtractJson parsing fallback, improved extraction prompts, retries and clearer parsing/logging
+- Initialize and teardown SmartAi and orchestrator in test setup/summary, and enable onToken streaming handlers for real-time output
+- Bump devDependencies: @push.rocks/smartagent to ^1.3.0 and @push.rocks/smartai to ^0.12.0
+
+## 2026-01-20 - 1.14.3 - fix(repo)
+no changes detected in the diff; no files modified and no release required
+
+- Diff contained no changes
+- No files were added, removed, or modified
+- No code, dependency, or documentation updates to release
+
+## 2026-01-19 - 1.14.2 - fix(readme)
+update README to document Nanonets-OCR2-3B (replaces Nanonets-OCR-s), adjust VRAM and context defaults, expand feature docs, and update examples/test command
+
+- Renamed Nanonets-OCR-s -> Nanonets-OCR2-3B throughout README and examples
+- Updated Nanonets VRAM guidance from ~10GB to ~12-16GB and documented 30K context
+- Changed documented MAX_MODEL_LEN default from 8192 to 30000
+- Updated example model identifiers (model strings and curl/example snippets) to nanonets/Nanonets-OCR2-3B
+- Added MiniCPM and Qwen feature bullets (multilingual, multi-image, flowchart support, expanded context notes)
+- Replaced README test command from ./test-images.sh to pnpm test
+
+## 2026-01-19 - 1.14.1 - fix(extraction)
+improve JSON extraction prompts and model options for invoice and bank statement tests
+
+- Refactor JSON extraction prompts to be sent after the document text and add explicit 'WHERE TO FIND DATA' and 'RULES' sections for clearer extraction guidance
+- Change chat message flow to: send document, assistant acknowledgement, then the JSON extraction prompt (avoids concatenating large prompts into one message)
+- Add model options (num_ctx: 32768, temperature: 0) to give larger context windows and deterministic JSON output
+- Simplify logging to avoid printing full prompt contents; log document and prompt lengths instead
+- Increase timeouts for large documents to 600000ms (10 minutes) where applicable
+
+## 2026-01-19 - 1.14.0 - feat(docker-images)
+add vLLM-based Nanonets-OCR2-3B image, Qwen3-VL Ollama image and refactor build/docs/tests to use new runtime/layout
+
+- Add new Dockerfiles for Nanonets (Dockerfile_nanonets_vllm_gpu_VRAM10GB), Qwen3 (Dockerfile_qwen3vl_ollama_gpu_VRAM20GB) and a clarified MiniCPM Ollama variant (Dockerfile_minicpm45v_ollama_gpu_VRAM9GB); remove older, redundant Dockerfiles.
+- Update build-images.sh to build the new image tags (minicpm45v, qwen3vl, nanonets-ocr) and adjust messaging/targets accordingly.
+- Documentation overhaul: readme.md and readme.hints.md updated to reflect vLLM vs Ollama runtimes, corrected ports/VRAM estimates, volume recommendations, and API endpoint details.
+- Tests updated to target the new model ID (nanonets/Nanonets-OCR2-3B), to process one page per batch, and to include a 10-minute AbortSignal timeout for OCR requests.
+- Added focused extraction test suites (test/test.invoices.extraction.ts and test/test.invoices.failed.ts) for faster iteration and debugging of invoice extraction.
+- Bump devDependencies: @git.zone/tsrun -> ^2.0.1 and @git.zone/tstest -> ^3.1.5.
+- Misc: test helper references and docker compose/test port mapping fixed (nanonets uses 8000), and various README sections cleaned and reorganized.
+
+## 2026-01-18 - 1.13.2 - fix(tests)
+stabilize OCR extraction tests and manage GPU containers
+
+- Add stopAllGpuContainers() and call it before starting GPU images to free GPU memory.
+- Remove PaddleOCR-VL image configs and associated ensure helpers from docker test helper to simplify images list.
+- Split invoice/bankstatement tests into two sequential stages: Stage 1 runs Nanonets OCR to produce markdown files, Stage 2 stops Nanonets and runs model extraction from saved markdown (avoids GPU contention).
+- Introduce temporary markdown directory handling and cleanup; add stopNanonets() and container running checks in tests.
+- Switch bank statement extraction model from qwen3:8b to gpt-oss:20b; add request timeout and improved logging/console output across tests.
+- Refactor extractWithConsensus and extraction functions to accept document identifiers, improve error messages and JSON extraction robustness.
+
+## 2026-01-18 - 1.13.1 - fix(image_support_files)
+remove PaddleOCR-VL server scripts from image_support_files
+
+- Deleted files: image_support_files/paddleocr_vl_full_server.py (approx. 636 lines) and image_support_files/paddleocr_vl_server.py (approx. 465 lines)
+- Cleanup/removal of legacy PaddleOCR-VL FastAPI server implementations — may affect users who relied on these local scripts
+
 ## 2026-01-18 - 1.13.0 - feat(tests)
 revamp tests and remove legacy Dockerfiles: adopt JSON/consensus workflows, switch MiniCPM model, and delete deprecated Docker/test variants
 

image_support_files/paddleocr_vl_full_server.py

@@ -1,636 +0,0 @@
-#!/usr/bin/env python3
-"""
-PaddleOCR-VL Full Pipeline API Server (Transformers backend)
-
-Provides REST API for document parsing using:
-- PP-DocLayoutV2 for layout detection
-- PaddleOCR-VL (transformers) for recognition
-- Structured JSON/Markdown output
-"""
-
-import os
-import io
-import re
-import base64
-import logging
-import tempfile
-import time
-import json
-from typing import Optional, List, Union
-from pathlib import Path
-
-from fastapi import FastAPI, HTTPException, UploadFile, File, Form
-from fastapi.responses import JSONResponse
-from pydantic import BaseModel
-from PIL import Image
-import torch
-
-# Configure logging
-logging.basicConfig(
-    level=logging.INFO,
-    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
-)
-logger = logging.getLogger(__name__)
-
-# Environment configuration
-SERVER_HOST = os.environ.get('SERVER_HOST', '0.0.0.0')
-SERVER_PORT = int(os.environ.get('SERVER_PORT', '8000'))
-MODEL_NAME = "PaddlePaddle/PaddleOCR-VL"
-
-# Device configuration
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-logger.info(f"Using device: {DEVICE}")
-
-# Task prompts
-TASK_PROMPTS = {
-    "ocr": "OCR:",
-    "table": "Table Recognition:",
-    "formula": "Formula Recognition:",
-    "chart": "Chart Recognition:",
-}
-
-# Initialize FastAPI app
-app = FastAPI(
-    title="PaddleOCR-VL Full Pipeline Server",
-    description="Document parsing with PP-DocLayoutV2 + PaddleOCR-VL (transformers)",
-    version="1.0.0"
-)
-
-# Global model instances
-vl_model = None
-vl_processor = None
-layout_model = None
-
-
-def load_vl_model():
-    """Load the PaddleOCR-VL model for element recognition"""
-    global vl_model, vl_processor
-
-    if vl_model is not None:
-        return
-
-    logger.info(f"Loading PaddleOCR-VL model: {MODEL_NAME}")
-    from transformers import AutoModelForCausalLM, AutoProcessor
-
-    vl_processor = AutoProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)
-
-    if DEVICE == "cuda":
-        vl_model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            trust_remote_code=True,
-            torch_dtype=torch.bfloat16,
-        ).to(DEVICE).eval()
-    else:
-        vl_model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            trust_remote_code=True,
-            torch_dtype=torch.float32,
-            low_cpu_mem_usage=True,
-        ).eval()
-
-    logger.info("PaddleOCR-VL model loaded successfully")
-
-
-def load_layout_model():
-    """Load the LayoutDetection model for layout detection"""
-    global layout_model
-
-    if layout_model is not None:
-        return
-
-    try:
-        logger.info("Loading LayoutDetection model (PP-DocLayout_plus-L)...")
-        from paddleocr import LayoutDetection
-
-        layout_model = LayoutDetection()
-        logger.info("LayoutDetection model loaded successfully")
-    except Exception as e:
-        logger.warning(f"Could not load LayoutDetection: {e}")
-        logger.info("Falling back to VL-only mode (no layout detection)")
-
-
-def recognize_element(image: Image.Image, task: str = "ocr") -> str:
-    """Recognize a single element using PaddleOCR-VL"""
-    load_vl_model()
-
-    prompt = TASK_PROMPTS.get(task, TASK_PROMPTS["ocr"])
-
-    messages = [
-        {
-            "role": "user",
-            "content": [
-                {"type": "image", "image": image},
-                {"type": "text", "text": prompt},
-            ]
-        }
-    ]
-
-    inputs = vl_processor.apply_chat_template(
-        messages,
-        tokenize=True,
-        add_generation_prompt=True,
-        return_dict=True,
-        return_tensors="pt"
-    )
-
-    if DEVICE == "cuda":
-        inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
-
-    with torch.inference_mode():
-        outputs = vl_model.generate(
-            **inputs,
-            max_new_tokens=4096,
-            do_sample=False,
-            use_cache=True
-        )
-
-    response = vl_processor.batch_decode(outputs, skip_special_tokens=True)[0]
-
-    # Extract only the assistant's response content
-    # The response format is: "User: <prompt>\nAssistant: <content>"
-    # We want to extract just the content after "Assistant:"
-    if "Assistant:" in response:
-        parts = response.split("Assistant:")
-        if len(parts) > 1:
-            response = parts[-1].strip()
-    elif "assistant:" in response.lower():
-        # Case-insensitive fallback
-        import re
-        match = re.split(r'[Aa]ssistant:', response)
-        if len(match) > 1:
-            response = match[-1].strip()
-
-    return response
-
-
-def detect_layout(image: Image.Image) -> List[dict]:
-    """Detect layout regions in the image"""
-    load_layout_model()
-
-    if layout_model is None:
-        # No layout model - return a single region covering the whole image
-        return [{
-            "type": "text",
-            "bbox": [0, 0, image.width, image.height],
-            "score": 1.0
-        }]
-
-    # Save image to temp file
-    with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp:
-        image.save(tmp.name, "PNG")
-        tmp_path = tmp.name
-
-    try:
-        results = layout_model.predict(tmp_path)
-        regions = []
-
-        for res in results:
-            # LayoutDetection returns boxes in 'boxes' key
-            for box in res.get("boxes", []):
-                coord = box.get("coordinate", [0, 0, image.width, image.height])
-                # Convert numpy floats to regular floats
-                bbox = [float(c) for c in coord]
-                regions.append({
-                    "type": box.get("label", "text"),
-                    "bbox": bbox,
-                    "score": float(box.get("score", 1.0))
-                })
-
-        # Sort regions by vertical position (top to bottom)
-        regions.sort(key=lambda r: r["bbox"][1])
-
-        return regions if regions else [{
-            "type": "text",
-            "bbox": [0, 0, image.width, image.height],
-            "score": 1.0
-        }]
-
-    finally:
-        os.unlink(tmp_path)
-
-
-def process_document(image: Image.Image) -> dict:
-    """Process a document through the full pipeline"""
-    logger.info(f"Processing document: {image.size}")
-
-    # Step 1: Detect layout
-    regions = detect_layout(image)
-    logger.info(f"Detected {len(regions)} layout regions")
-
-    # Step 2: Recognize each region
-    blocks = []
-    for i, region in enumerate(regions):
-        region_type = region["type"].lower()
-        bbox = region["bbox"]
-
-        # Crop region from image
-        x1, y1, x2, y2 = [int(c) for c in bbox]
-        region_image = image.crop((x1, y1, x2, y2))
-
-        # Determine task based on region type
-        if "table" in region_type:
-            task = "table"
-        elif "formula" in region_type or "math" in region_type:
-            task = "formula"
-        elif "chart" in region_type or "figure" in region_type:
-            task = "chart"
-        else:
-            task = "ocr"
-
-        # Recognize the region
-        try:
-            content = recognize_element(region_image, task)
-            blocks.append({
-                "index": i,
-                "type": region_type,
-                "bbox": bbox,
-                "content": content,
-                "task": task
-            })
-            logger.info(f"  Region {i} ({region_type}): {len(content)} chars")
-        except Exception as e:
-            logger.error(f"  Region {i} error: {e}")
-            blocks.append({
-                "index": i,
-                "type": region_type,
-                "bbox": bbox,
-                "content": "",
-                "error": str(e)
-            })
-
-    return {"blocks": blocks, "image_size": list(image.size)}
-
-
-def result_to_markdown(result: dict) -> str:
-    """Convert result to Markdown format with structural hints for LLM processing.
-
-    Adds positional and type-based formatting to help downstream LLMs
-    understand document structure:
-    - Tables are marked with **[TABLE]** prefix
-    - Header zone content (top 15%) is bolded
-    - Footer zone content (bottom 15%) is separated with horizontal rule
-    - Titles are formatted as # headers
-    - Figures/charts are marked with *[Figure: ...]*
-    """
-    lines = []
-    image_height = result.get("image_size", [0, 1000])[1]
-
-    for block in result.get("blocks", []):
-        block_type = block.get("type", "text").lower()
-        content = block.get("content", "").strip()
-        bbox = block.get("bbox", [])
-
-        if not content:
-            continue
-
-        # Determine position zone (top 15%, middle, bottom 15%)
-        y_pos = bbox[1] if bbox and len(bbox) > 1 else 0
-        y_end = bbox[3] if bbox and len(bbox) > 3 else y_pos
-        is_header_zone = y_pos < image_height * 0.15
-        is_footer_zone = y_end > image_height * 0.85
-
-        # Format based on type and position
-        if "table" in block_type:
-            lines.append(f"\n**[TABLE]**\n{content}\n")
-        elif "title" in block_type:
-            lines.append(f"# {content}")
-        elif "formula" in block_type or "math" in block_type:
-            lines.append(f"\n$$\n{content}\n$$\n")
-        elif "figure" in block_type or "chart" in block_type:
-            lines.append(f"*[Figure: {content}]*")
-        elif is_header_zone:
-            lines.append(f"**{content}**")
-        elif is_footer_zone:
-            lines.append(f"---\n{content}")
-        else:
-            lines.append(content)
-
-    return "\n\n".join(lines)
-
-
-def parse_markdown_table(content: str) -> str:
-    """Convert table content to HTML table.
-
-    Handles:
-    - PaddleOCR-VL format: <fcel>cell<lcel>cell<nl> (detected by <fcel> tags)
-    - Pipe-delimited tables: | Header | Header |
-    - Separator rows: |---|---|
-    - Returns HTML <table> structure
-    """
-    content_stripped = content.strip()
-
-    # Check for PaddleOCR-VL table format (<fcel>, <lcel>, <ecel>, <nl>)
-    if '<fcel>' in content_stripped or '<nl>' in content_stripped:
-        return parse_paddleocr_table(content_stripped)
-
-    lines = content_stripped.split('\n')
-    if not lines:
-        return f'<pre>{content}</pre>'
-
-    # Check if it looks like a markdown table
-    if not any('|' in line for line in lines):
-        return f'<pre>{content}</pre>'
-
-    html_rows = []
-    is_header = True
-
-    for line in lines:
-        line = line.strip()
-        if not line or line.startswith('|') == False and '|' not in line:
-            continue
-
-        # Skip separator rows (|---|---|)
-        if re.match(r'^[\|\s\-:]+$', line):
-            is_header = False
-            continue
-
-        # Parse cells
-        cells = [c.strip() for c in line.split('|')]
-        cells = [c for c in cells if c]  # Remove empty from edges
-
-        if is_header:
-            row = '<tr>' + ''.join(f'<th>{c}</th>' for c in cells) + '</tr>'
-            html_rows.append(f'<thead>{row}</thead>')
-            is_header = False
-        else:
-            row = '<tr>' + ''.join(f'<td>{c}</td>' for c in cells) + '</tr>'
-            html_rows.append(row)
-
-    if html_rows:
-        # Wrap body rows in tbody
-        header = html_rows[0] if '<thead>' in html_rows[0] else ''
-        body_rows = [r for r in html_rows if '<thead>' not in r]
-        body = f'<tbody>{"".join(body_rows)}</tbody>' if body_rows else ''
-        return f'<table>{header}{body}</table>'
-
-    return f'<pre>{content}</pre>'
-
-
-def parse_paddleocr_table(content: str) -> str:
-    """Convert PaddleOCR-VL table format to HTML table.
-
-    PaddleOCR-VL uses:
-    - <fcel> = first cell in a row
-    - <lcel> = subsequent cells
-    - <ecel> = empty cell
-    - <nl> = row separator (newline)
-
-    Example input:
-    <fcel>Header1<lcel>Header2<nl><fcel>Value1<lcel>Value2<nl>
-    """
-    # Split into rows by <nl>
-    rows_raw = re.split(r'<nl>', content)
-    html_rows = []
-    is_first_row = True
-
-    for row_content in rows_raw:
-        row_content = row_content.strip()
-        if not row_content:
-            continue
-
-        # Extract cells: split by <fcel>, <lcel>, or <ecel>
-        # Each cell is the text between these markers
-        cells = []
-
-        # Pattern to match cell markers and capture content
-        # Content is everything between markers
-        parts = re.split(r'<fcel>|<lcel>|<ecel>', row_content)
-        for part in parts:
-            part = part.strip()
-            if part:
-                cells.append(part)
-
-        if not cells:
-            continue
-
-        # First row is header
-        if is_first_row:
-            row_html = '<tr>' + ''.join(f'<th>{c}</th>' for c in cells) + '</tr>'
-            html_rows.append(f'<thead>{row_html}</thead>')
-            is_first_row = False
-        else:
-            row_html = '<tr>' + ''.join(f'<td>{c}</td>' for c in cells) + '</tr>'
-            html_rows.append(row_html)
-
-    if html_rows:
-        header = html_rows[0] if '<thead>' in html_rows[0] else ''
-        body_rows = [r for r in html_rows if '<thead>' not in r]
-        body = f'<tbody>{"".join(body_rows)}</tbody>' if body_rows else ''
-        return f'<table>{header}{body}</table>'
-
-    return f'<pre>{content}</pre>'
-
-
-def result_to_html(result: dict) -> str:
-    """Convert result to semantic HTML for optimal LLM processing.
-
-    Uses semantic HTML5 tags with position metadata as data-* attributes.
-    Markdown tables are converted to proper HTML <table> tags for
-    unambiguous parsing by downstream LLMs.
-    """
-    parts = []
-    image_height = result.get("image_size", [0, 1000])[1]
-
-    parts.append('<!DOCTYPE html><html><body>')
-
-    for block in result.get("blocks", []):
-        block_type = block.get("type", "text").lower()
-        content = block.get("content", "").strip()
-        bbox = block.get("bbox", [])
-
-        if not content:
-            continue
-
-        # Position metadata
-        y_pos = bbox[1] / image_height if bbox and len(bbox) > 1 else 0
-        data_attrs = f'data-type="{block_type}" data-y="{y_pos:.2f}"'
-
-        # Format based on type
-        if "table" in block_type:
-            table_html = parse_markdown_table(content)
-            parts.append(f'<section {data_attrs} class="table-region">{table_html}</section>')
-        elif "title" in block_type:
-            parts.append(f'<h1 {data_attrs}>{content}</h1>')
-        elif "formula" in block_type or "math" in block_type:
-            parts.append(f'<div {data_attrs} class="formula"><code>{content}</code></div>')
-        elif "figure" in block_type or "chart" in block_type:
-            parts.append(f'<figure {data_attrs}><figcaption>{content}</figcaption></figure>')
-        elif y_pos < 0.15:
-            parts.append(f'<header {data_attrs}><strong>{content}</strong></header>')
-        elif y_pos > 0.85:
-            parts.append(f'<footer {data_attrs}>{content}</footer>')
-        else:
-            parts.append(f'<p {data_attrs}>{content}</p>')
-
-    parts.append('</body></html>')
-    return '\n'.join(parts)
-
-
-# Request/Response models
-class ParseRequest(BaseModel):
-    image: str  # base64 encoded image
-    output_format: Optional[str] = "json"
-
-
-class ParseResponse(BaseModel):
-    success: bool
-    format: str
-    result: Union[dict, str]
-    processing_time: float
-    error: Optional[str] = None
-
-
-def decode_image(image_source: str) -> Image.Image:
-    """Decode image from base64 or data URL"""
-    if image_source.startswith("data:"):
-        header, data = image_source.split(",", 1)
-        image_data = base64.b64decode(data)
-    else:
-        image_data = base64.b64decode(image_source)
-
-    return Image.open(io.BytesIO(image_data)).convert("RGB")
-
-
-@app.on_event("startup")
-async def startup_event():
-    """Pre-load models on startup"""
-    logger.info("Starting PaddleOCR-VL Full Pipeline Server...")
-    try:
-        load_vl_model()
-        load_layout_model()
-        logger.info("Models loaded successfully")
-    except Exception as e:
-        logger.error(f"Failed to pre-load models: {e}")
-
-
-@app.get("/health")
-async def health_check():
-    """Health check endpoint"""
-    return {
-        "status": "healthy" if vl_model is not None else "loading",
-        "service": "PaddleOCR-VL Full Pipeline (Transformers)",
-        "device": DEVICE,
-        "vl_model_loaded": vl_model is not None,
-        "layout_model_loaded": layout_model is not None
-    }
-
-
-@app.get("/formats")
-async def supported_formats():
-    """List supported output formats"""
-    return {
-        "output_formats": ["json", "markdown", "html"],
-        "image_formats": ["PNG", "JPEG", "WebP", "BMP", "GIF", "TIFF"],
-        "capabilities": [
-            "Layout detection (PP-DocLayoutV2)",
-            "Text recognition (OCR)",
-            "Table recognition",
-            "Formula recognition (LaTeX)",
-            "Chart recognition",
-            "Multi-language support (109 languages)"
-        ]
-    }
-
-
-@app.post("/parse", response_model=ParseResponse)
-async def parse_document_endpoint(request: ParseRequest):
-    """Parse a document image and return structured output"""
-    try:
-        start_time = time.time()
-
-        image = decode_image(request.image)
-        result = process_document(image)
-
-        if request.output_format == "markdown":
-            markdown = result_to_markdown(result)
-            output = {"markdown": markdown}
-        elif request.output_format == "html":
-            html = result_to_html(result)
-            output = {"html": html}
-        else:
-            output = result
-
-        elapsed = time.time() - start_time
-        logger.info(f"Processing complete in {elapsed:.2f}s")
-
-        return ParseResponse(
-            success=True,
-            format=request.output_format,
-            result=output,
-            processing_time=elapsed
-        )
-
-    except Exception as e:
-        logger.error(f"Error processing document: {e}", exc_info=True)
-        return ParseResponse(
-            success=False,
-            format=request.output_format,
-            result={},
-            processing_time=0,
-            error=str(e)
-        )
-
-
-@app.post("/v1/chat/completions")
-async def chat_completions(request: dict):
-    """OpenAI-compatible chat completions endpoint"""
-    try:
-        messages = request.get("messages", [])
-        output_format = request.get("output_format", "json")
-
-        # Find user message with image
-        image = None
-        for msg in reversed(messages):
-            if msg.get("role") == "user":
-                content = msg.get("content", [])
-                if isinstance(content, list):
-                    for item in content:
-                        if item.get("type") == "image_url":
-                            url = item.get("image_url", {}).get("url", "")
-                            image = decode_image(url)
-                            break
-                break
-
-        if image is None:
-            raise HTTPException(status_code=400, detail="No image provided")
-
-        start_time = time.time()
-        result = process_document(image)
-
-        if output_format == "markdown":
-            content = result_to_markdown(result)
-        elif output_format == "html":
-            content = result_to_html(result)
-        else:
-            content = json.dumps(result, ensure_ascii=False, indent=2)
-
-        elapsed = time.time() - start_time
-
-        return {
-            "id": f"chatcmpl-{int(time.time()*1000)}",
-            "object": "chat.completion",
-            "created": int(time.time()),
-            "model": "paddleocr-vl-full",
-            "choices": [{
-                "index": 0,
-                "message": {"role": "assistant", "content": content},
-                "finish_reason": "stop"
-            }],
-            "usage": {
-                "prompt_tokens": 100,
-                "completion_tokens": len(content) // 4,
-                "total_tokens": 100 + len(content) // 4
-            },
-            "processing_time": elapsed
-        }
-
-    except HTTPException:
-        raise
-    except Exception as e:
-        logger.error(f"Error in chat completions: {e}", exc_info=True)
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-if __name__ == "__main__":
-    import uvicorn
-    uvicorn.run(app, host=SERVER_HOST, port=SERVER_PORT)

image_support_files/paddleocr_vl_server.py

@@ -1,465 +0,0 @@
-#!/usr/bin/env python3
-"""
-PaddleOCR-VL FastAPI Server (CPU variant)
-Provides OpenAI-compatible REST API for document parsing using PaddleOCR-VL
-"""
-
-import os
-import io
-import base64
-import logging
-import time
-from typing import Optional, List, Any, Dict, Union
-
-from fastapi import FastAPI, HTTPException
-from fastapi.responses import JSONResponse
-from pydantic import BaseModel
-import torch
-from PIL import Image
-
-# Configure logging
-logging.basicConfig(
-    level=logging.INFO,
-    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
-)
-logger = logging.getLogger(__name__)
-
-# Environment configuration
-SERVER_HOST = os.environ.get('SERVER_HOST', '0.0.0.0')
-SERVER_PORT = int(os.environ.get('SERVER_PORT', '8000'))
-MODEL_NAME = os.environ.get('MODEL_NAME', 'PaddlePaddle/PaddleOCR-VL')
-
-# Device configuration
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-logger.info(f"Using device: {DEVICE}")
-
-# Task prompts for PaddleOCR-VL
-TASK_PROMPTS = {
-    "ocr": "OCR:",
-    "table": "Table Recognition:",
-    "formula": "Formula Recognition:",
-    "chart": "Chart Recognition:",
-}
-
-# Initialize FastAPI app
-app = FastAPI(
-    title="PaddleOCR-VL Server",
-    description="OpenAI-compatible REST API for document parsing using PaddleOCR-VL",
-    version="1.0.0"
-)
-
-# Global model instances
-model = None
-processor = None
-
-
-# Request/Response models (OpenAI-compatible)
-class ImageUrl(BaseModel):
-    url: str
-
-
-class ContentItem(BaseModel):
-    type: str
-    text: Optional[str] = None
-    image_url: Optional[ImageUrl] = None
-
-
-class Message(BaseModel):
-    role: str
-    content: Union[str, List[ContentItem]]
-
-
-class ChatCompletionRequest(BaseModel):
-    model: str = "paddleocr-vl"
-    messages: List[Message]
-    temperature: Optional[float] = 0.0
-    max_tokens: Optional[int] = 4096
-
-
-class Choice(BaseModel):
-    index: int
-    message: Message
-    finish_reason: str
-
-
-class Usage(BaseModel):
-    prompt_tokens: int
-    completion_tokens: int
-    total_tokens: int
-
-
-class ChatCompletionResponse(BaseModel):
-    id: str
-    object: str = "chat.completion"
-    created: int
-    model: str
-    choices: List[Choice]
-    usage: Usage
-
-
-class HealthResponse(BaseModel):
-    status: str
-    model: str
-    device: str
-
-
-def load_model():
-    """Load the PaddleOCR-VL model and processor"""
-    global model, processor
-
-    if model is not None:
-        return
-
-    logger.info(f"Loading PaddleOCR-VL model: {MODEL_NAME}")
-
-    from transformers import AutoModelForCausalLM, AutoProcessor
-
-    # Load processor
-    processor = AutoProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)
-
-    # Load model with appropriate settings for CPU/GPU
-    if DEVICE == "cuda":
-        model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            trust_remote_code=True,
-            torch_dtype=torch.bfloat16,
-        ).to(DEVICE).eval()
-    else:
-        # CPU mode - use float32 for compatibility
-        model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            trust_remote_code=True,
-            torch_dtype=torch.float32,
-            low_cpu_mem_usage=True,
-        ).eval()
-
-    logger.info("PaddleOCR-VL model loaded successfully")
-
-
-def optimize_image_resolution(image: Image.Image, max_size: int = 2048, min_size: int = 1080) -> Image.Image:
-    """
-    Optimize image resolution for PaddleOCR-VL.
-
-    Best results are achieved with images in the 1080p-2K range.
-    - Images larger than max_size are scaled down
-    - Very small images are scaled up to min_size
-    """
-    width, height = image.size
-    max_dim = max(width, height)
-    min_dim = min(width, height)
-
-    # Scale down if too large (4K+ images often miss text)
-    if max_dim > max_size:
-        scale = max_size / max_dim
-        new_width = int(width * scale)
-        new_height = int(height * scale)
-        logger.info(f"Scaling down image from {width}x{height} to {new_width}x{new_height}")
-        image = image.resize((new_width, new_height), Image.Resampling.LANCZOS)
-    # Scale up if too small
-    elif max_dim < min_size and min_dim < min_size:
-        scale = min_size / max_dim
-        new_width = int(width * scale)
-        new_height = int(height * scale)
-        logger.info(f"Scaling up image from {width}x{height} to {new_width}x{new_height}")
-        image = image.resize((new_width, new_height), Image.Resampling.LANCZOS)
-    else:
-        logger.info(f"Image size {width}x{height} is optimal, no scaling needed")
-
-    return image
-
-
-def decode_image(image_source: str, optimize: bool = True) -> Image.Image:
-    """
-    Decode image from various sources.
-
-    Supported formats:
-    - Base64 data URL: data:image/png;base64,... or data:image/jpeg;base64,...
-    - HTTP/HTTPS URL: https://example.com/image.png
-    - Raw base64 string
-    - Local file path
-
-    Supported image types: PNG, JPEG, WebP, BMP, GIF, TIFF
-    """
-    image = None
-
-    if image_source.startswith("data:"):
-        # Base64 encoded image with MIME type header
-        # Supports: data:image/png;base64,... data:image/jpeg;base64,... etc.
-        header, data = image_source.split(",", 1)
-        image_data = base64.b64decode(data)
-        image = Image.open(io.BytesIO(image_data)).convert("RGB")
-        logger.debug(f"Decoded base64 image with header: {header}")
-    elif image_source.startswith("http://") or image_source.startswith("https://"):
-        # URL - fetch image
-        import httpx
-        response = httpx.get(image_source, timeout=30.0)
-        response.raise_for_status()
-        image = Image.open(io.BytesIO(response.content)).convert("RGB")
-        logger.debug(f"Fetched image from URL: {image_source[:50]}...")
-    else:
-        # Assume it's a file path or raw base64
-        try:
-            image_data = base64.b64decode(image_source)
-            image = Image.open(io.BytesIO(image_data)).convert("RGB")
-            logger.debug("Decoded raw base64 image")
-        except:
-            # Try as file path
-            image = Image.open(image_source).convert("RGB")
-            logger.debug(f"Loaded image from file: {image_source}")
-
-    # Optimize resolution for best OCR results
-    if optimize:
-        image = optimize_image_resolution(image)
-
-    return image
-
-
-def extract_image_and_text(content: Union[str, List[ContentItem]]) -> tuple:
-    """Extract image and text prompt from message content"""
-    if isinstance(content, str):
-        return None, content
-
-    image = None
-    text = ""
-
-    for item in content:
-        if item.type == "image_url" and item.image_url:
-            image = decode_image(item.image_url.url)
-        elif item.type == "text" and item.text:
-            text = item.text
-
-    return image, text
-
-
-def generate_response(image: Image.Image, prompt: str, max_tokens: int = 4096) -> str:
-    """Generate response using PaddleOCR-VL"""
-    load_model()
-
-    messages = [
-        {
-            "role": "user",
-            "content": [
-                {"type": "image", "image": image},
-                {"type": "text", "text": prompt},
-            ]
-        }
-    ]
-
-    inputs = processor.apply_chat_template(
-        messages,
-        tokenize=True,
-        add_generation_prompt=True,
-        return_dict=True,
-        return_tensors="pt"
-    )
-
-    if DEVICE == "cuda":
-        inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
-
-    with torch.inference_mode():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=max_tokens,
-            do_sample=False,
-            use_cache=True
-        )
-
-    response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
-
-    # Extract the assistant's response (after the prompt)
-    if "assistant" in response.lower():
-        parts = response.split("assistant")
-        if len(parts) > 1:
-            response = parts[-1].strip()
-
-    return response
-
-
-@app.on_event("startup")
-async def startup_event():
-    """Pre-load the model on startup"""
-    logger.info("Pre-loading PaddleOCR-VL model...")
-    try:
-        load_model()
-        logger.info("Model pre-loaded successfully")
-    except Exception as e:
-        logger.error(f"Failed to pre-load model: {e}")
-        # Don't fail startup - model will be loaded on first request
-
-
-@app.get("/health", response_model=HealthResponse)
-async def health_check():
-    """Health check endpoint"""
-    return HealthResponse(
-        status="healthy" if model is not None else "loading",
-        model=MODEL_NAME,
-        device=DEVICE
-    )
-
-
-@app.get("/formats")
-async def supported_formats():
-    """List supported image formats and input methods"""
-    return {
-        "image_formats": {
-            "supported": ["PNG", "JPEG", "WebP", "BMP", "GIF", "TIFF"],
-            "recommended": ["PNG", "JPEG"],
-            "mime_types": [
-                "image/png",
-                "image/jpeg",
-                "image/webp",
-                "image/bmp",
-                "image/gif",
-                "image/tiff"
-            ]
-        },
-        "input_methods": {
-            "base64_data_url": {
-                "description": "Base64 encoded image with MIME type header",
-                "example": "data:image/png;base64,iVBORw0KGgo..."
-            },
-            "http_url": {
-                "description": "Direct HTTP/HTTPS URL to image",
-                "example": "https://example.com/image.png"
-            },
-            "raw_base64": {
-                "description": "Raw base64 string without header",
-                "example": "iVBORw0KGgo..."
-            }
-        },
-        "resolution": {
-            "optimal_range": "1080p to 2K (1080-2048 pixels on longest side)",
-            "auto_scaling": True,
-            "note": "Images are automatically scaled to optimal range. 4K+ images are scaled down for better accuracy."
-        },
-        "task_prompts": TASK_PROMPTS
-    }
-
-
-@app.get("/v1/models")
-async def list_models():
-    """List available models (OpenAI-compatible)"""
-    return {
-        "object": "list",
-        "data": [
-            {
-                "id": "paddleocr-vl",
-                "object": "model",
-                "created": int(time.time()),
-                "owned_by": "paddlepaddle"
-            }
-        ]
-    }
-
-
-@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
-async def chat_completions(request: ChatCompletionRequest):
-    """
-    OpenAI-compatible chat completions endpoint for PaddleOCR-VL
-
-    Supports tasks:
-    - "OCR:" - Text recognition
-    - "Table Recognition:" - Table extraction
-    - "Formula Recognition:" - Formula extraction
-    - "Chart Recognition:" - Chart extraction
-    """
-    try:
-        # Get the last user message
-        user_message = None
-        for msg in reversed(request.messages):
-            if msg.role == "user":
-                user_message = msg
-                break
-
-        if not user_message:
-            raise HTTPException(status_code=400, detail="No user message found")
-
-        # Extract image and prompt
-        image, prompt = extract_image_and_text(user_message.content)
-
-        if image is None:
-            raise HTTPException(status_code=400, detail="No image provided in message")
-
-        # Default to OCR if no specific prompt
-        if not prompt or prompt.strip() == "":
-            prompt = "OCR:"
-
-        logger.info(f"Processing request with prompt: {prompt[:50]}...")
-
-        # Generate response
-        start_time = time.time()
-        response_text = generate_response(image, prompt, request.max_tokens or 4096)
-        elapsed = time.time() - start_time
-
-        logger.info(f"Generated response in {elapsed:.2f}s ({len(response_text)} chars)")
-
-        # Build OpenAI-compatible response
-        return ChatCompletionResponse(
-            id=f"chatcmpl-{int(time.time()*1000)}",
-            created=int(time.time()),
-            model=request.model,
-            choices=[
-                Choice(
-                    index=0,
-                    message=Message(role="assistant", content=response_text),
-                    finish_reason="stop"
-                )
-            ],
-            usage=Usage(
-                prompt_tokens=100,  # Approximate
-                completion_tokens=len(response_text) // 4,
-                total_tokens=100 + len(response_text) // 4
-            )
-        )
-
-    except HTTPException:
-        raise
-    except Exception as e:
-        logger.error(f"Error processing request: {e}")
-        raise HTTPException(status_code=500, detail=str(e))
-
-
-# Legacy endpoint for compatibility with old PaddleOCR API
-class LegacyOCRRequest(BaseModel):
-    image: str
-    task: Optional[str] = "ocr"
-
-
-class LegacyOCRResponse(BaseModel):
-    success: bool
-    result: str
-    task: str
-    error: Optional[str] = None
-
-
-@app.post("/ocr", response_model=LegacyOCRResponse)
-async def legacy_ocr(request: LegacyOCRRequest):
-    """
-    Legacy OCR endpoint for backwards compatibility
-
-    Tasks: ocr, table, formula, chart
-    """
-    try:
-        image = decode_image(request.image)
-        prompt = TASK_PROMPTS.get(request.task, TASK_PROMPTS["ocr"])
-
-        result = generate_response(image, prompt)
-
-        return LegacyOCRResponse(
-            success=True,
-            result=result,
-            task=request.task
-        )
-    except Exception as e:
-        logger.error(f"Legacy OCR error: {e}")
-        return LegacyOCRResponse(
-            success=False,
-            result="",
-            task=request.task,
-            error=str(e)
-        )
-
-
-if __name__ == "__main__":
-    import uvicorn
-    uvicorn.run(app, host=SERVER_HOST, port=SERVER_PORT)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@host.today/ht-docker-ai",
-  "version": "1.13.0",
+  "version": "1.15.1",
   "type": "module",
   "private": false,
   "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",
@@ -13,8 +13,10 @@
     "test": "tstest test/ --verbose"
   },
   "devDependencies": {
-    "@git.zone/tsrun": "^1.3.3",
-    "@git.zone/tstest": "^1.0.90"
+    "@git.zone/tsrun": "^2.0.1",
+    "@git.zone/tstest": "^3.1.5",
+    "@push.rocks/smartagent": "^1.5.4",
+    "@push.rocks/smartai": "^0.13.2"
   },
   "repository": {
     "type": "git",

readme.hints.md

@@ -2,12 +2,18 @@
 
 ## Architecture
 
-This project uses **Ollama** as the runtime framework for serving AI models. This provides:
+This project uses **Ollama** and **vLLM** as runtime frameworks for serving AI models:
 
+### Ollama-based Images (MiniCPM-V, Qwen3-VL)
 - Automatic model download and caching
 - Unified REST API (compatible with OpenAI format)
 - Built-in quantization support
-- GPU/CPU auto-detection
+- GPU auto-detection
+
+### vLLM-based Images (Nanonets-OCR)
+- High-performance inference server
+- OpenAI-compatible API
+- Optimized for VLM workloads
 
 ## Model Details
 
@@ -24,18 +30,24 @@ This project uses **Ollama** as the runtime framework for serving AI models. Thi
 |------|---------------|
 | Full precision (bf16) | 18GB |
 | int4 quantized | 9GB |
-| GGUF (CPU) | 8GB RAM |
 
 ## Container Startup Flow
 
+### Ollama-based containers
 1. `docker-entrypoint.sh` starts Ollama server in background
 2. Waits for server to be ready
 3. Checks if model already exists in volume
 4. Pulls model if not present
 5. Keeps container running
 
+### vLLM-based containers
+1. vLLM server starts with model auto-download
+2. Health check endpoint available at `/health`
+3. OpenAI-compatible API at `/v1/chat/completions`
+
 ## Volume Persistence
 
+### Ollama volumes
 Mount `/root/.ollama` to persist downloaded models:
 
 ```bash
@@ -44,9 +56,16 @@ Mount `/root/.ollama` to persist downloaded models:
 
 Without this volume, the model will be re-downloaded on each container start (~5GB download).
 
+### vLLM/HuggingFace volumes
+Mount `/root/.cache/huggingface` for model caching:
+
+```bash
+-v hf-cache:/root/.cache/huggingface
+```
+
 ## API Endpoints
 
-All endpoints follow the Ollama API specification:
+### Ollama API (MiniCPM-V, Qwen3-VL)
 
 | Endpoint | Method | Description |
 |----------|--------|-------------|
@@ -56,113 +75,137 @@ All endpoints follow the Ollama API specification:
 | `/api/pull` | POST | Pull a model |
 | `/api/show` | POST | Show model info |
 
-## GPU Detection
+### vLLM API (Nanonets-OCR)
 
-The GPU variant uses Ollama's automatic GPU detection. For CPU-only mode, we set:
-
-```dockerfile
-ENV CUDA_VISIBLE_DEVICES=""
-```
-
-This forces Ollama to use CPU inference even if GPU is available.
+| Endpoint | Method | Description |
+|----------|--------|-------------|
+| `/health` | GET | Health check |
+| `/v1/models` | GET | List available models |
+| `/v1/chat/completions` | POST | OpenAI-compatible chat completions |
 
 ## Health Checks
 
-Both variants include Docker health checks:
+All containers include Docker health checks:
 
 ```dockerfile
 HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
     CMD curl -f http://localhost:11434/api/tags || exit 1
 ```
 
-CPU variant has longer `start-period` (120s) due to slower startup.
+---
 
-## PaddleOCR-VL (Recommended)
+## Nanonets-OCR-s
 
 ### Overview
 
-PaddleOCR-VL is a 0.9B parameter Vision-Language Model specifically optimized for document parsing. It replaces the older PP-Structure approach with native VLM understanding.
+Nanonets-OCR-s is a Qwen2.5-VL-3B model fine-tuned specifically for document OCR tasks. It outputs structured markdown with semantic tags.
 
-**Key advantages over PP-Structure:**
-- Native table understanding (no HTML parsing needed)
-- 109 language support
-- Better handling of complex multi-row tables
-- Structured Markdown/JSON output
+**Key features:**
+- Based on Qwen2.5-VL-3B (~4B parameters)
+- Fine-tuned for document OCR
+- Outputs markdown with semantic HTML tags
+- ~10GB VRAM
 
 ### Docker Images
 
 | Tag | Description |
 |-----|-------------|
-| `paddleocr-vl` | GPU variant using vLLM (recommended) |
-| `paddleocr-vl-cpu` | CPU variant using transformers |
+| `nanonets-ocr` | GPU variant using vLLM (OpenAI-compatible API) |
 
-### API Endpoints (OpenAI-compatible)
+### API Endpoints (OpenAI-compatible via vLLM)
 
 | Endpoint | Method | Description |
 |----------|--------|-------------|
-| `/health` | GET | Health check with model info |
+| `/health` | GET | Health check |
 | `/v1/models` | GET | List available models |
 | `/v1/chat/completions` | POST | OpenAI-compatible chat completions |
-| `/ocr` | POST | Legacy OCR endpoint |
 
 ### Request/Response Format
 
 **POST /v1/chat/completions (OpenAI-compatible)**
 ```json
 {
-  "model": "paddleocr-vl",
+  "model": "nanonets/Nanonets-OCR-s",
   "messages": [
     {
       "role": "user",
       "content": [
         {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
-        {"type": "text", "text": "Table Recognition:"}
+        {"type": "text", "text": "Extract the text from the above document..."}
       ]
     }
   ],
   "temperature": 0.0,
-  "max_tokens": 8192
+  "max_tokens": 4096
 }
 ```
 
-**Task Prompts:**
-- `"OCR:"` - Text recognition
-- `"Table Recognition:"` - Table extraction (returns markdown)
-- `"Formula Recognition:"` - Formula extraction
-- `"Chart Recognition:"` - Chart extraction
+### Nanonets OCR Prompt
 
-**Response**
-```json
-{
-  "id": "chatcmpl-...",
-  "object": "chat.completion",
-  "choices": [
-    {
-      "index": 0,
-      "message": {
-        "role": "assistant",
-        "content": "| Date | Description | Amount |\n|---|---|---|\n| 2021-06-01 | GITLAB INC | -119.96 |"
-      },
-      "finish_reason": "stop"
-    }
-  ]
-}
+The model is designed to work with a specific prompt format:
+```
+Extract the text from the above document as if you were reading it naturally.
+Return the tables in html format.
+Return the equations in LaTeX representation.
+If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
+Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
+Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.
 ```
-
-### Environment Variables
-
-| Variable | Default | Description |
-|----------|---------|-------------|
-| `MODEL_NAME` | `PaddlePaddle/PaddleOCR-VL` | Model to load |
-| `HOST` | `0.0.0.0` | Server host |
-| `PORT` | `8000` | Server port |
-| `MAX_BATCHED_TOKENS` | `16384` | vLLM max batch tokens |
-| `GPU_MEMORY_UTILIZATION` | `0.9` | GPU memory usage (0-1) |
 
 ### Performance
 
-- **GPU (vLLM)**: ~2-5 seconds per page
-- **CPU**: ~30-60 seconds per page
+- **GPU (vLLM)**: ~3-8 seconds per page
+- **VRAM usage**: ~10GB
+
+### Two-Stage Pipeline (Nanonets + Qwen3)
+
+The Nanonets tests use a two-stage pipeline:
+1. **Stage 1**: Nanonets-OCR-s converts images to markdown (via vLLM on port 8000)
+2. **Stage 2**: Qwen3 8B extracts structured JSON from markdown (via Ollama on port 11434)
+
+**GPU Limitation**: Both vLLM and Ollama require significant GPU memory. On a single GPU system:
+- Running both simultaneously causes memory contention
+- For single GPU: Run services sequentially (stop Nanonets before Qwen3)
+- For multi-GPU: Assign each service to a different GPU
+
+**Sequential Execution**:
+```bash
+# Step 1: Run Nanonets OCR (converts to markdown)
+docker start nanonets-test
+# ... perform OCR ...
+docker stop nanonets-test
+
+# Step 2: Run Qwen3 extraction (from markdown)
+docker start minicpm-test
+# ... extract JSON ...
+```
+
+---
+
+## Multi-Pass Extraction Strategy
+
+The bank statement extraction uses a dual-VLM consensus approach:
+
+### Architecture: Dual-VLM Consensus
+
+| VLM | Model | Purpose |
+|-----|-------|---------|
+| **MiniCPM-V 4.5** | 8B params | Primary visual extraction |
+| **Nanonets-OCR-s** | ~4B params | Document OCR with semantic output |
+
+### Extraction Strategy
+
+1. **Pass 1**: MiniCPM-V visual extraction (images → JSON)
+2. **Pass 2**: Nanonets-OCR semantic extraction (images → markdown → JSON)
+3. **Consensus**: If Pass 1 == Pass 2 → Done (fast path)
+4. **Pass 3+**: MiniCPM-V visual if no consensus
+
+### Why Dual-VLM Works
+
+- **Different architectures**: Two independent models cross-check each other
+- **Specialized strengths**: Nanonets-OCR-s optimized for document structure, MiniCPM-V for general vision
+- **No structure loss**: Both VLMs see the original images directly
+- **Fast consensus**: Most documents complete in 2 passes when VLMs agree
 
 ---
 
@@ -170,7 +213,7 @@ PaddleOCR-VL is a 0.9B parameter Vision-Language Model specifically optimized fo
 
 To add a new model variant:
 
-1. Create `Dockerfile_<modelname>`
+1. Create `Dockerfile_<modelname>_<runtime>_<hardware>_VRAM<size>`
 2. Set `MODEL_NAME` environment variable
 3. Update `build-images.sh` with new build target
 4. Add documentation to `readme.md`
@@ -188,8 +231,8 @@ The model download is ~5GB and may take several minutes.
 
 ### Out of memory
 
-- GPU: Use int4 quantized version or add more VRAM
-- CPU: Increase container memory limit: `--memory=16g`
+- GPU: Use a lighter model variant or upgrade VRAM
+- Add more GPU memory: Consider multi-GPU setup
 
 ### API not responding
 
@@ -207,41 +250,6 @@ npmci docker build
 npmci docker push code.foss.global
 ```
 
-## Multi-Pass Extraction Strategy
-
-The bank statement extraction uses a dual-VLM consensus approach:
-
-### Architecture: Dual-VLM Consensus
-
-| VLM | Model | Purpose |
-|-----|-------|---------|
-| **MiniCPM-V 4.5** | 8B params | Primary visual extraction |
-| **PaddleOCR-VL** | 0.9B params | Table-specialized extraction |
-
-### Extraction Strategy
-
-1. **Pass 1**: MiniCPM-V visual extraction (images → JSON)
-2. **Pass 2**: PaddleOCR-VL table recognition (images → markdown → JSON)
-3. **Consensus**: If Pass 1 == Pass 2 → Done (fast path)
-4. **Pass 3+**: MiniCPM-V visual if no consensus
-
-### Why Dual-VLM Works
-
-- **Different architectures**: Two independent models cross-check each other
-- **Specialized strengths**: PaddleOCR-VL optimized for tables, MiniCPM-V for general vision
-- **No structure loss**: Both VLMs see the original images directly
-- **Fast consensus**: Most documents complete in 2 passes when VLMs agree
-
-### Comparison vs Old PP-Structure Approach
-
-| Approach | Bank Statement Result | Issue |
-|----------|----------------------|-------|
-| MiniCPM-V Visual | 28 transactions ✓ | - |
-| PP-Structure HTML + Visual | 13 transactions ✗ | HTML merged rows incorrectly |
-| PaddleOCR-VL Table | 28 transactions ✓ | Native table understanding |
-
-**Key insight**: PP-Structure's HTML output loses structure for complex tables. PaddleOCR-VL's native VLM approach maintains table integrity.
-
 ---
 
 ## Related Resources
@@ -249,3 +257,4 @@ The bank statement extraction uses a dual-VLM consensus approach:
 - [Ollama Documentation](https://ollama.ai/docs)
 - [MiniCPM-V GitHub](https://github.com/OpenBMB/MiniCPM-V)
 - [Ollama API Reference](https://github.com/ollama/ollama/blob/main/docs/api.md)
+- [Nanonets-OCR-s on HuggingFace](https://huggingface.co/nanonets/Nanonets-OCR-s)

readme.md

@@ -1,40 +1,52 @@
 # @host.today/ht-docker-ai 🚀
 
-Production-ready Docker images for state-of-the-art AI Vision-Language Models. Run powerful multimodal AI locally with GPU acceleration or CPU fallback—no cloud API keys required.
+Production-ready Docker images for state-of-the-art AI Vision-Language Models. Run powerful multimodal AI locally with GPU acceleration—**no cloud API keys required**.
+
+> 🔥 **Three VLMs, one registry.** From high-performance document OCR to GPT-4o-level vision understanding—pick the right tool for your task.
 
 ## Issue Reporting and Security
 
 For reporting bugs, issues, or security vulnerabilities, please visit [community.foss.global/](https://community.foss.global/). This is the central community hub for all issue reporting. Developers who sign and comply with our contribution agreement and go through identification can also get a [code.foss.global/](https://code.foss.global/) account to submit Pull Requests directly.
 
+---
+
 ## 🎯 What's Included
 
-| Model | Parameters | Best For | API |
-|-------|-----------|----------|-----|
-| **MiniCPM-V 4.5** | 8B | General vision understanding, image analysis, multi-image | Ollama-compatible |
-| **PaddleOCR-VL** | 0.9B | Document parsing, table extraction, OCR | OpenAI-compatible |
+| Model | Parameters | Best For | API | Port | VRAM |
+|-------|-----------|----------|-----|------|------|
+| **MiniCPM-V 4.5** | 8B | General vision understanding, multi-image analysis | Ollama-compatible | 11434 | ~9GB |
+| **Nanonets-OCR2-3B** | ~3B | Document OCR with semantic markdown, LaTeX, flowcharts | OpenAI-compatible | 8000 | ~12-16GB |
+| **Qwen3-VL-30B** | 30B (A3B) | Advanced visual agents, code generation from images | Ollama-compatible | 11434 | ~20GB |
 
-## 📦 Available Images
+---
+
+## 📦 Quick Reference: All Available Images
 
 ```
 code.foss.global/host.today/ht-docker-ai:<tag>
 ```
 
-| Tag | Model | Hardware | Port |
-|-----|-------|----------|------|
-| `minicpm45v` / `latest` | MiniCPM-V 4.5 | NVIDIA GPU (9-18GB VRAM) | 11434 |
-| `minicpm45v-cpu` | MiniCPM-V 4.5 | CPU only (8GB+ RAM) | 11434 |
-| `paddleocr-vl` / `paddleocr-vl-gpu` | PaddleOCR-VL | NVIDIA GPU | 8000 |
-| `paddleocr-vl-cpu` | PaddleOCR-VL | CPU only | 8000 |
+| Tag | Model | Runtime | Port | VRAM |
+|-----|-------|---------|------|------|
+| `minicpm45v` / `latest` | MiniCPM-V 4.5 | Ollama | 11434 | ~9GB |
+| `nanonets-ocr` | Nanonets-OCR2-3B | vLLM | 8000 | ~12-16GB |
+| `qwen3vl` | Qwen3-VL-30B-A3B | Ollama | 11434 | ~20GB |
 
 ---
 
 ## 🖼️ MiniCPM-V 4.5
 
-A GPT-4o level multimodal LLM from OpenBMB—handles image understanding, OCR, multi-image analysis, and visual reasoning across 30+ languages.
+A GPT-4o level multimodal LLM from OpenBMB—handles image understanding, OCR, multi-image analysis, and visual reasoning across **30+ languages**.
+
+### ✨ Key Features
+
+- 🌍 **Multilingual:** 30+ languages supported
+- 🖼️ **Multi-image:** Analyze multiple images in one request
+- 📊 **Versatile:** Charts, documents, photos, diagrams
+- ⚡ **Efficient:** Runs on consumer GPUs (9GB VRAM)
 
 ### Quick Start
 
-**GPU (Recommended):**
 ```bash
 docker run -d \
   --name minicpm \
@@ -44,15 +56,6 @@ docker run -d \
   code.foss.global/host.today/ht-docker-ai:minicpm45v
 ```
 
-**CPU Only:**
-```bash
-docker run -d \
-  --name minicpm \
-  -p 11434:11434 \
-  -v ollama-data:/root/.ollama \
-  code.foss.global/host.today/ht-docker-ai:minicpm45v-cpu
-```
-
 > 💡 **Pro tip:** Mount the volume to persist downloaded models (~5GB). Without it, models re-download on every container start.
 
 ### API Examples
@@ -85,110 +88,131 @@ curl http://localhost:11434/api/chat -d '{
 
 ### Hardware Requirements
 
-| Variant | VRAM/RAM | Notes |
-|---------|----------|-------|
-| GPU (int4 quantized) | 9GB VRAM | Recommended for most use cases |
-| GPU (full precision) | 18GB VRAM | Maximum quality |
-| CPU (GGUF) | 8GB+ RAM | Slower but accessible |
+| Mode | VRAM Required |
+|------|---------------|
+| int4 quantized | ~9GB |
+| Full precision (bf16) | ~18GB |
 
 ---
 
-## 📄 PaddleOCR-VL
+## 🔍 Nanonets-OCR2-3B
 
-A specialized 0.9B Vision-Language Model optimized for document parsing. Native support for tables, formulas, charts, and text extraction in 109 languages.
+The **latest Nanonets document OCR model** (October 2025 release)—based on Qwen2.5-VL-3B, fine-tuned specifically for document extraction with significant improvements over the original OCR-s.
+
+### ✨ Key Features
+
+- 📝 **Semantic output:** Tables → HTML, equations → LaTeX, flowcharts → structured markup
+- 🌍 **Multilingual:** Inherits Qwen's broad language support
+- 📄 **30K context:** Handle large, multi-page documents
+- 🔌 **OpenAI-compatible:** Drop-in replacement for existing pipelines
+- 🎯 **Improved accuracy:** Better semantic tagging and LaTeX equation extraction vs. OCR-s
 
 ### Quick Start
 
-**GPU:**
 ```bash
 docker run -d \
-  --name paddleocr \
+  --name nanonets \
   --gpus all \
   -p 8000:8000 \
   -v hf-cache:/root/.cache/huggingface \
-  code.foss.global/host.today/ht-docker-ai:paddleocr-vl
+  code.foss.global/host.today/ht-docker-ai:nanonets-ocr
 ```
 
-**CPU:**
-```bash
-docker run -d \
-  --name paddleocr \
-  -p 8000:8000 \
-  -v hf-cache:/root/.cache/huggingface \
-  code.foss.global/host.today/ht-docker-ai:paddleocr-vl-cpu
-```
-
-### OpenAI-Compatible API
-
-PaddleOCR-VL exposes a fully OpenAI-compatible `/v1/chat/completions` endpoint:
+### API Usage
 
 ```bash
 curl http://localhost:8000/v1/chat/completions \
   -H "Content-Type: application/json" \
   -d '{
-    "model": "paddleocr-vl",
+    "model": "nanonets/Nanonets-OCR2-3B",
     "messages": [{
       "role": "user",
       "content": [
         {"type": "image_url", "image_url": {"url": "data:image/png;base64,<base64>"}},
-        {"type": "text", "text": "Table Recognition:"}
+        {"type": "text", "text": "Extract the text from the above document as if you were reading it naturally. Return the tables in html format. Return the equations in LaTeX representation."}
       ]
     }],
-    "max_tokens": 8192
+    "temperature": 0.0,
+    "max_tokens": 4096
   }'
 ```
 
-### Task Prompts
+### Output Format
 
-| Prompt | Output | Use Case |
-|--------|--------|----------|
-| `OCR:` | Plain text | General text extraction |
-| `Table Recognition:` | Markdown table | Invoices, bank statements, spreadsheets |
-| `Formula Recognition:` | LaTeX | Math equations, scientific notation |
-| `Chart Recognition:` | Description | Graphs and visualizations |
+Nanonets-OCR2-3B returns markdown with semantic tags:
 
-### API Endpoints
+| Element | Output Format |
+|---------|---------------|
+| Tables | `<table>...</table>` (HTML) |
+| Equations | `$...$` (LaTeX) |
+| Images | `<img>description</img>` |
+| Watermarks | `<watermark>OFFICIAL COPY</watermark>` |
+| Page numbers | `<page_number>14</page_number>` |
+| Flowcharts | Structured markup |
 
-| Endpoint | Method | Description |
-|----------|--------|-------------|
-| `/health` | GET | Health check with model/device info |
-| `/formats` | GET | Supported image formats and input methods |
-| `/v1/models` | GET | List available models |
-| `/v1/chat/completions` | POST | OpenAI-compatible chat completions |
-| `/ocr` | POST | Legacy OCR endpoint |
+### Hardware Requirements
 
-### Image Input Methods
+| Config | VRAM |
+|--------|------|
+| 30K context (default) | ~12-16GB |
+| Speed | ~3-8 seconds per page |
 
-PaddleOCR-VL accepts images in multiple formats:
+---
 
-```javascript
-// Base64 data URL
-"data:image/png;base64,iVBORw0KGgo..."
+## 🧠 Qwen3-VL-30B-A3B
 
-// HTTP URL
-"https://example.com/document.png"
+The **most powerful** Qwen vision model—30B parameters with 3B active (MoE architecture). Handles complex visual reasoning, code generation from screenshots, and visual agent capabilities.
 
-// Raw base64
-"iVBORw0KGgo..."
+### ✨ Key Features
+
+- 🚀 **256K context** (expandable to 1M tokens!)
+- 🤖 **Visual agent capabilities** — can plan and execute multi-step tasks
+- 💻 **Code generation from images** — screenshot → working code
+- 🎯 **State-of-the-art** visual reasoning
+
+### Quick Start
+
+```bash
+docker run -d \
+  --name qwen3vl \
+  --gpus all \
+  -p 11434:11434 \
+  -v ollama-data:/root/.ollama \
+  code.foss.global/host.today/ht-docker-ai:qwen3vl
 ```
 
-**Supported formats:** PNG, JPEG, WebP, BMP, GIF, TIFF
+Then pull the model (one-time, ~20GB):
+```bash
+docker exec qwen3vl ollama pull qwen3-vl:30b-a3b
+```
 
-**Optimal resolution:** 1080p–2K. Images are automatically scaled for best results.
+### API Usage
 
-### Performance
+```bash
+curl http://localhost:11434/api/chat -d '{
+  "model": "qwen3-vl:30b-a3b",
+  "messages": [{
+    "role": "user",
+    "content": "Analyze this screenshot and write the code to recreate this UI",
+    "images": ["<base64-encoded-image>"]
+  }]
+}'
+```
 
-| Mode | Speed per Page |
-|------|----------------|
-| GPU (CUDA) | 2–5 seconds |
-| CPU | 30–60 seconds |
+### Hardware Requirements
+
+| Requirement | Value |
+|-------------|-------|
+| VRAM | ~20GB (Q4_K_M quantization) |
+| Context | 256K tokens default |
 
 ---
 
 ## 🐳 Docker Compose
 
+Run multiple VLMs together for maximum flexibility:
+
 ```yaml
-version: '3.8'
 services:
   # General vision tasks
   minicpm:
@@ -206,9 +230,9 @@ services:
               capabilities: [gpu]
     restart: unless-stopped
 
-  # Document parsing / OCR
-  paddleocr:
-    image: code.foss.global/host.today/ht-docker-ai:paddleocr-vl
+  # Document OCR with semantic output
+  nanonets:
+    image: code.foss.global/host.today/ht-docker-ai:nanonets-ocr
     ports:
       - "8000:8000"
     volumes:
@@ -231,7 +255,7 @@ volumes:
 
 ## ⚙️ Environment Variables
 
-### MiniCPM-V 4.5
+### MiniCPM-V 4.5 & Qwen3-VL (Ollama-based)
 
 | Variable | Default | Description |
 |----------|---------|-------------|
@@ -239,13 +263,47 @@ volumes:
 | `OLLAMA_HOST` | `0.0.0.0` | API bind address |
 | `OLLAMA_ORIGINS` | `*` | Allowed CORS origins |
 
-### PaddleOCR-VL
+### Nanonets-OCR (vLLM-based)
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `MODEL_NAME` | `PaddlePaddle/PaddleOCR-VL` | HuggingFace model ID |
-| `SERVER_HOST` | `0.0.0.0` | API bind address |
-| `SERVER_PORT` | `8000` | API port |
+| `MODEL_NAME` | `nanonets/Nanonets-OCR2-3B` | HuggingFace model ID |
+| `HOST` | `0.0.0.0` | API bind address |
+| `PORT` | `8000` | API port |
+| `MAX_MODEL_LEN` | `30000` | Maximum sequence length |
+| `GPU_MEMORY_UTILIZATION` | `0.9` | GPU memory usage (0-1) |
+
+---
+
+## 🏗️ Architecture Notes
+
+### Dual-VLM Consensus Strategy
+
+For production document extraction, consider using multiple models together:
+
+1. **Pass 1:** MiniCPM-V visual extraction (images → JSON)
+2. **Pass 2:** Nanonets-OCR semantic extraction (images → markdown → JSON)
+3. **Consensus:** If results match → Done (fast path)
+4. **Pass 3+:** Additional visual passes if needed
+
+This dual-VLM approach catches extraction errors that single models miss.
+
+### Why Multi-Model Works
+
+- **Different architectures:** Independent models cross-validate each other
+- **Specialized strengths:** Nanonets-OCR2-3B excels at document structure; MiniCPM-V handles general vision
+- **Native processing:** All VLMs see original images—no intermediate structure loss
+
+### Model Selection Guide
+
+| Task | Recommended Model |
+|------|-------------------|
+| General image understanding | MiniCPM-V 4.5 |
+| Document OCR with structure preservation | Nanonets-OCR2-3B |
+| Complex visual reasoning / code generation | Qwen3-VL-30B |
+| Multi-image analysis | MiniCPM-V 4.5 |
+| Visual agent tasks | Qwen3-VL-30B |
+| Large documents (30K+ tokens) | Nanonets-OCR2-3B |
 
 ---
 
@@ -260,42 +318,21 @@ cd ht-docker-ai
 ./build-images.sh
 
 # Run tests
-./test-images.sh
+pnpm test
 ```
 
 ---
 
-## 🏗️ Architecture Notes
-
-### Dual-VLM Consensus Strategy
-
-For production document extraction, consider using both models together:
-
-1. **Pass 1:** MiniCPM-V visual extraction (images → JSON)
-2. **Pass 2:** PaddleOCR-VL table recognition (images → markdown → JSON)
-3. **Consensus:** If results match → Done (fast path)
-4. **Pass 3+:** Additional visual passes if needed
-
-This dual-VLM approach catches extraction errors that single models miss.
-
-### Why This Works
-
-- **Different architectures:** Two independent models cross-validate each other
-- **Specialized strengths:** PaddleOCR-VL excels at tables; MiniCPM-V handles general vision
-- **Native processing:** Both VLMs see original images—no intermediate HTML/structure loss
-
----
-
 ## 🔍 Troubleshooting
 
 ### Model download hangs
 ```bash
 docker logs -f <container-name>
 ```
-Model downloads can take several minutes (~5GB for MiniCPM-V).
+Model downloads can take several minutes (~5GB for MiniCPM-V, ~20GB for Qwen3-VL).
 
 ### Out of memory
-- **GPU:** Use the CPU variant or upgrade VRAM
+- **GPU:** Use a lighter model variant or upgrade VRAM
 - **CPU:** Increase container memory: `--memory=16g`
 
 ### API not responding
@@ -315,6 +352,13 @@ sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 ```
 
+### GPU Memory Contention (Multi-Model)
+
+When running multiple VLMs on a single GPU:
+- vLLM and Ollama both need significant GPU memory
+- **Single GPU:** Run services sequentially (stop one before starting another)
+- **Multi-GPU:** Assign each service to a different GPU via `CUDA_VISIBLE_DEVICES`
+
 ---
 
 ## License and Legal Information

test/helpers/docker.ts

@@ -2,11 +2,8 @@ import { execSync } from 'child_process';
 
 // Project container names (only manage these)
 const PROJECT_CONTAINERS = [
-  'paddleocr-vl-test',
-  'paddleocr-vl-gpu-test',
-  'paddleocr-vl-cpu-test',
-  'paddleocr-vl-full-test',
   'minicpm-test',
+  'nanonets-test',
 ];
 
 // Image configurations
@@ -23,30 +20,6 @@ export interface IImageConfig {
 }
 
 export const IMAGES = {
-  paddleocrVlGpu: {
-    name: 'paddleocr-vl-gpu',
-    dockerfile: 'Dockerfile_paddleocr_vl_gpu',
-    buildContext: '.',
-    containerName: 'paddleocr-vl-test',
-    ports: ['8000:8000'],
-    volumes: ['ht-huggingface-cache:/root/.cache/huggingface'],
-    gpus: true,
-    healthEndpoint: 'http://localhost:8000/health',
-    healthTimeout: 300000, // 5 minutes for model loading
-  } as IImageConfig,
-
-  paddleocrVlCpu: {
-    name: 'paddleocr-vl-cpu',
-    dockerfile: 'Dockerfile_paddleocr_vl_cpu',
-    buildContext: '.',
-    containerName: 'paddleocr-vl-test',
-    ports: ['8000:8000'],
-    volumes: ['ht-huggingface-cache:/root/.cache/huggingface'],
-    gpus: false,
-    healthEndpoint: 'http://localhost:8000/health',
-    healthTimeout: 300000,
-  } as IImageConfig,
-
   minicpm: {
     name: 'minicpm45v',
     dockerfile: 'Dockerfile_minicpm45v_gpu',
@@ -59,20 +32,17 @@ export const IMAGES = {
     healthTimeout: 120000,
   } as IImageConfig,
 
-  // Full PaddleOCR-VL pipeline with PP-DocLayoutV2 + structured JSON output
-  paddleocrVlFull: {
-    name: 'paddleocr-vl-full',
-    dockerfile: 'Dockerfile_paddleocr_vl_full',
+  // Nanonets-OCR2-3B - Document OCR optimized VLM (Qwen2.5-VL-3B fine-tuned, Oct 2025)
+  nanonetsOcr: {
+    name: 'nanonets-ocr',
+    dockerfile: 'Dockerfile_nanonets_vllm_gpu_VRAM10GB',
     buildContext: '.',
-    containerName: 'paddleocr-vl-full-test',
+    containerName: 'nanonets-test',
     ports: ['8000:8000'],
-    volumes: [
-      'ht-huggingface-cache:/root/.cache/huggingface',
-      'ht-paddleocr-cache:/root/.paddleocr',
-    ],
+    volumes: ['ht-huggingface-cache:/root/.cache/huggingface'],
     gpus: true,
     healthEndpoint: 'http://localhost:8000/health',
-    healthTimeout: 600000, // 10 minutes for model loading (vLLM + PP-DocLayoutV2)
+    healthTimeout: 300000, // 5 minutes for model loading
   } as IImageConfig,
 };
 
@@ -126,7 +96,7 @@ export function removeContainer(containerName: string): void {
 }
 
 /**
- * Stop all project containers that conflict with the required one
+ * Stop all project containers that conflict with the required one (port-based)
  */
 export function stopConflictingContainers(requiredContainer: string, requiredPort: string): void {
   // Stop project containers using the same port
@@ -144,6 +114,24 @@ export function stopConflictingContainers(requiredContainer: string, requiredPor
   }
 }
 
+/**
+ * Stop all GPU-consuming project containers (for GPU memory management)
+ * This ensures GPU memory is freed before starting a new GPU service
+ */
+export function stopAllGpuContainers(exceptContainer?: string): void {
+  for (const container of PROJECT_CONTAINERS) {
+    if (container === exceptContainer) continue;
+
+    if (isContainerRunning(container)) {
+      console.log(`[Docker] Stopping GPU container: ${container}`);
+      exec(`docker stop ${container}`, true);
+      // Give the GPU a moment to free memory
+    }
+  }
+  // Brief pause to allow GPU memory to be released
+  execSync('sleep 2');
+}
+
 /**
  * Build a Docker image
  */
@@ -220,6 +208,11 @@ export async function ensureService(config: IImageConfig): Promise<boolean> {
     buildImage(config);
   }
 
+  // For GPU services, stop ALL other GPU containers to free GPU memory
+  if (config.gpus) {
+    stopAllGpuContainers(config.containerName);
+  }
+
   // Stop conflicting containers on the same port
   const mainPort = config.ports[0];
   stopConflictingContainers(config.containerName, mainPort);
@@ -240,21 +233,7 @@ export async function ensureService(config: IImageConfig): Promise<boolean> {
 }
 
 /**
- * Ensure PaddleOCR-VL GPU service is running
- */
-export async function ensurePaddleOcrVlGpu(): Promise<boolean> {
-  return ensureService(IMAGES.paddleocrVlGpu);
-}
-
-/**
- * Ensure PaddleOCR-VL CPU service is running
- */
-export async function ensurePaddleOcrVlCpu(): Promise<boolean> {
-  return ensureService(IMAGES.paddleocrVlCpu);
-}
-
-/**
- * Ensure MiniCPM service is running
+ * Ensure MiniCPM service is running (Ollama with GPU)
  */
 export async function ensureMiniCpm(): Promise<boolean> {
   return ensureService(IMAGES.minicpm);
@@ -272,30 +251,6 @@ export function isGpuAvailable(): boolean {
   }
 }
 
-/**
- * Ensure PaddleOCR-VL service (auto-detect GPU/CPU)
- */
-export async function ensurePaddleOcrVl(): Promise<boolean> {
-  if (isGpuAvailable()) {
-    console.log('[Docker] GPU detected, using GPU image');
-    return ensurePaddleOcrVlGpu();
-  } else {
-    console.log('[Docker] No GPU detected, using CPU image');
-    return ensurePaddleOcrVlCpu();
-  }
-}
-
-/**
- * Ensure PaddleOCR-VL Full Pipeline service (PP-DocLayoutV2 + structured output)
- * This is the recommended service for production use - outputs structured JSON/Markdown
- */
-export async function ensurePaddleOcrVlFull(): Promise<boolean> {
-  if (!isGpuAvailable()) {
-    console.log('[Docker] WARNING: Full pipeline requires GPU, but none detected');
-  }
-  return ensureService(IMAGES.paddleocrVlFull);
-}
-
 /**
  * Ensure an Ollama model is pulled and available
  * Uses the MiniCPM container (which runs Ollama) to pull the model
@@ -383,3 +338,14 @@ export async function ensureQwen3Vl(): Promise<boolean> {
   // Then ensure Qwen3-VL 8B is pulled
   return ensureOllamaModel('qwen3-vl:8b');
 }
+
+/**
+ * Ensure Nanonets-OCR2-3B service is running (via vLLM)
+ * Document OCR optimized VLM based on Qwen2.5-VL-3B (Oct 2025 release)
+ */
+export async function ensureNanonetsOcr(): Promise<boolean> {
+  if (!isGpuAvailable()) {
+    console.log('[Docker] WARNING: Nanonets-OCR2-3B requires GPU, but none detected');
+  }
+  return ensureService(IMAGES.nanonetsOcr);
+}

test/test.bankstatements.minicpm.ts

@@ -1,9 +1,10 @@
 /**
- * Bank statement extraction using MiniCPM-V (visual extraction)
+ * Bank statement extraction using MiniCPM-V via smartagent DualAgentOrchestrator
  *
- * JSON per-page approach:
- * 1. Ask for structured JSON of all transactions per page
- * 2. Consensus: extract twice, compare, retry if mismatch
+ * Uses vision-capable orchestrator with JsonValidatorTool for self-validation:
+ * 1. Process each page with the orchestrator
+ * 2. Driver extracts transactions and validates JSON before completing
+ * 3. Streaming output during extraction
  */
 import { tap, expect } from '@git.zone/tstest/tapbundle';
 import * as fs from 'fs';
@@ -11,6 +12,8 @@ import * as path from 'path';
 import { execSync } from 'child_process';
 import * as os from 'os';
 import { ensureMiniCpm } from './helpers/docker.js';
+import { SmartAi } from '@push.rocks/smartai';
+import { DualAgentOrchestrator, JsonValidatorTool } from '@push.rocks/smartagent';
 
 const OLLAMA_URL = 'http://localhost:11434';
 const MODEL = 'openbmb/minicpm-v4.5:q8_0';
@@ -21,21 +24,9 @@ interface ITransaction {
   amount: number;
 }
 
-const JSON_PROMPT = `Extract ALL transactions from this bank statement page as a JSON array.
-
-IMPORTANT RULES:
-1. Each transaction has: date, description/counterparty, and an amount
-2. Amount is NEGATIVE for money going OUT (debits, payments, withdrawals)
-3. Amount is POSITIVE for money coming IN (credits, deposits, refunds)
-4. Date format: YYYY-MM-DD
-5. Do NOT include: opening balance, closing balance, subtotals, headers, or summary rows
-6. Only include actual transactions with a specific date and amount
-
-Return ONLY this JSON format, no explanation:
-[
-  {"date": "2021-06-01", "counterparty": "COMPANY NAME", "amount": -25.99},
-  {"date": "2021-06-02", "counterparty": "DEPOSIT FROM", "amount": 100.00}
-]`;
+// SmartAi instance and orchestrator (initialized in setup)
+let smartAi: SmartAi;
+let orchestrator: DualAgentOrchestrator;
 
 /**
  * Convert PDF to PNG images using ImageMagick
@@ -65,206 +56,31 @@ function convertPdfToImages(pdfPath: string): string[] {
   }
 }
 
-/**
- * Query for JSON extraction
- */
-async function queryJson(image: string, queryId: string): Promise<string> {
-  console.log(`      [${queryId}] Sending request to ${MODEL}...`);
-  const startTime = Date.now();
+const EXTRACTION_PROMPT = `Extract ALL transactions from this bank statement page as a JSON array.
 
-  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      model: MODEL,
-      messages: [{
-        role: 'user',
-        content: JSON_PROMPT,
-        images: [image],
-      }],
-      stream: false,
-      options: {
-        num_predict: 4000,
-        temperature: 0.1,
-      },
-    }),
-  });
+IMPORTANT RULES:
+1. Each transaction has: date, counterparty (description), and an amount
+2. Amount is NEGATIVE for money going OUT (debits, payments, withdrawals)
+3. Amount is POSITIVE for money coming IN (credits, deposits, refunds)
+4. Date format: YYYY-MM-DD
+5. Do NOT include: opening balance, closing balance, subtotals, headers, or summary rows
+6. Only include actual transactions with a specific date and amount
 
-  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+Before completing, validate your JSON output:
 
-  if (!response.ok) {
-    console.log(`      [${queryId}] ERROR: ${response.status} (${elapsed}s)`);
-    throw new Error(`Ollama API error: ${response.status}`);
-  }
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON_ARRAY_HERE"}</params>
+</tool_call>
 
-  const data = await response.json();
-  const content = (data.message?.content || '').trim();
-  console.log(`      [${queryId}] Response received (${elapsed}s, ${content.length} chars)`);
-  return content;
-}
+Output format (must be a valid JSON array):
+[
+  {"date": "2021-06-01", "counterparty": "COMPANY NAME", "amount": -25.99},
+  {"date": "2021-06-02", "counterparty": "DEPOSIT FROM", "amount": 100.00}
+]
 
-/**
- * Sanitize JSON string - fix common issues from vision model output
- */
-function sanitizeJson(jsonStr: string): string {
-  let s = jsonStr;
-
-  // Fix +number (e.g., +93.80 -> 93.80) - JSON doesn't allow + prefix
-  // Handle various whitespace patterns
-  s = s.replace(/"amount"\s*:\s*\+/g, '"amount": ');
-  s = s.replace(/:\s*\+(\d)/g, ': $1');
-
-  // Fix European number format with thousands separator (e.g., 1.000.00 -> 1000.00)
-  // Pattern: "amount": X.XXX.XX where X.XXX is thousands and .XX is decimal
-  s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3.$4');
-  // Also handle larger numbers like 10.000.00
-  s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3$4.$5');
-
-  // Fix trailing commas before ] or }
-  s = s.replace(/,\s*([}\]])/g, '$1');
-
-  // Fix unescaped newlines inside strings (replace with space)
-  s = s.replace(/"([^"\\]*)\n([^"]*)"/g, '"$1 $2"');
-
-  // Fix unescaped tabs inside strings
-  s = s.replace(/"([^"\\]*)\t([^"]*)"/g, '"$1 $2"');
-
-  // Fix unescaped backslashes (but not already escaped ones)
-  s = s.replace(/\\(?!["\\/bfnrtu])/g, '\\\\');
-
-  // Fix common issues with counterparty names containing special chars
-  s = s.replace(/"counterparty":\s*"([^"]*)'([^"]*)"/g, '"counterparty": "$1$2"');
-
-  // Remove control characters except newlines (which we handle above)
-  s = s.replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, ' ');
-
-  return s;
-}
-
-/**
- * Parse JSON response into transactions
- */
-function parseJsonResponse(response: string, queryId: string): ITransaction[] {
-  console.log(`      [${queryId}] Parsing response...`);
-
-  // Try to find JSON in markdown code block
-  const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
-  let jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
-
-  if (codeBlockMatch) {
-    console.log(`      [${queryId}] Found JSON in code block`);
-  }
-
-  // Sanitize JSON (fix +number issue)
-  jsonStr = sanitizeJson(jsonStr);
-
-  try {
-    const parsed = JSON.parse(jsonStr);
-    if (Array.isArray(parsed)) {
-      const txs = parsed.map(tx => ({
-        date: String(tx.date || ''),
-        counterparty: String(tx.counterparty || tx.description || ''),
-        amount: parseAmount(tx.amount),
-      }));
-      console.log(`      [${queryId}] Parsed ${txs.length} transactions (direct)`);
-      return txs;
-    }
-    console.log(`      [${queryId}] Parsed JSON is not an array`);
-  } catch (e) {
-    const errMsg = (e as Error).message;
-    console.log(`      [${queryId}] Direct parse failed: ${errMsg}`);
-
-    // Log problematic section with context
-    const posMatch = errMsg.match(/position (\d+)/);
-    if (posMatch) {
-      const pos = parseInt(posMatch[1]);
-      const start = Math.max(0, pos - 40);
-      const end = Math.min(jsonStr.length, pos + 40);
-      const context = jsonStr.substring(start, end);
-      const marker = ' '.repeat(pos - start) + '^';
-      console.log(`      [${queryId}] Context around error position ${pos}:`);
-      console.log(`      [${queryId}] ...${context}...`);
-      console.log(`      [${queryId}]    ${marker}`);
-    }
-
-    // Try to find JSON array pattern
-    const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
-    if (arrayMatch) {
-      console.log(`      [${queryId}] Found array pattern, trying to parse...`);
-      const sanitizedArray = sanitizeJson(arrayMatch[0]);
-      try {
-        const parsed = JSON.parse(sanitizedArray);
-        if (Array.isArray(parsed)) {
-          const txs = parsed.map(tx => ({
-            date: String(tx.date || ''),
-            counterparty: String(tx.counterparty || tx.description || ''),
-            amount: parseAmount(tx.amount),
-          }));
-          console.log(`      [${queryId}] Parsed ${txs.length} transactions (array match)`);
-          return txs;
-        }
-      } catch (e2) {
-        const errMsg2 = (e2 as Error).message;
-        console.log(`      [${queryId}] Array parse failed: ${errMsg2}`);
-        const posMatch2 = errMsg2.match(/position (\d+)/);
-        if (posMatch2) {
-          const pos2 = parseInt(posMatch2[1]);
-          console.log(`      [${queryId}] Context around error: ...${sanitizedArray.substring(Math.max(0, pos2 - 30), pos2 + 30)}...`);
-        }
-
-        // Try to extract individual objects from the malformed array
-        console.log(`      [${queryId}] Attempting object-by-object extraction...`);
-        const extracted = extractTransactionsFromMalformedJson(sanitizedArray, queryId);
-        if (extracted.length > 0) {
-          console.log(`      [${queryId}] Recovered ${extracted.length} transactions via object extraction`);
-          return extracted;
-        }
-      }
-    } else {
-      console.log(`      [${queryId}] No array pattern found in response`);
-      console.log(`      [${queryId}] Raw response preview: ${response.substring(0, 200)}...`);
-    }
-  }
-
-  console.log(`      [${queryId}] PARSE FAILED - returning empty array`);
-  return [];
-}
-
-/**
- * Extract transactions from malformed JSON by parsing objects individually
- */
-function extractTransactionsFromMalformedJson(jsonStr: string, queryId: string): ITransaction[] {
-  const transactions: ITransaction[] = [];
-
-  // Match individual transaction objects
-  const objectPattern = /\{\s*"date"\s*:\s*"([^"]+)"\s*,\s*"counterparty"\s*:\s*"([^"]+)"\s*,\s*"amount"\s*:\s*([+-]?\d+\.?\d*)\s*\}/g;
-  let match;
-
-  while ((match = objectPattern.exec(jsonStr)) !== null) {
-    transactions.push({
-      date: match[1],
-      counterparty: match[2],
-      amount: parseFloat(match[3]),
-    });
-  }
-
-  // Also try with different field orders (amount before counterparty, etc.)
-  if (transactions.length === 0) {
-    const altPattern = /\{\s*"date"\s*:\s*"([^"]+)"[^}]*"amount"\s*:\s*([+-]?\d+\.?\d*)[^}]*\}/g;
-    while ((match = altPattern.exec(jsonStr)) !== null) {
-      // Try to extract counterparty from the match
-      const counterpartyMatch = match[0].match(/"counterparty"\s*:\s*"([^"]+)"/);
-      const descMatch = match[0].match(/"description"\s*:\s*"([^"]+)"/);
-      transactions.push({
-        date: match[1],
-        counterparty: counterpartyMatch?.[1] || descMatch?.[1] || 'UNKNOWN',
-        amount: parseFloat(match[2]),
-      });
-    }
-  }
-
-  return transactions;
-}
+Only complete after validation passes. Output the final JSON array in <task_complete> tags.`;
 
 /**
  * Parse amount from various formats
@@ -284,102 +100,101 @@ function parseAmount(value: unknown): number {
 }
 
 /**
- * Compare two transaction arrays for consensus
+ * Extract JSON from response (handles markdown code blocks and task_complete tags)
  */
-function transactionArraysMatch(a: ITransaction[], b: ITransaction[]): boolean {
-  if (a.length !== b.length) return false;
-
-  for (let i = 0; i < a.length; i++) {
-    const dateMatch = a[i].date === b[i].date;
-    const amountMatch = Math.abs(a[i].amount - b[i].amount) < 0.01;
-    if (!dateMatch || !amountMatch) return false;
-  }
-
-  return true;
-}
-
-/**
- * Compare two transaction arrays and log differences
- */
-function compareAndLogDifferences(txs1: ITransaction[], txs2: ITransaction[], pageNum: number): void {
-  if (txs1.length !== txs2.length) {
-    console.log(`    [Page ${pageNum}] Length mismatch: Q1=${txs1.length}, Q2=${txs2.length}`);
-    return;
-  }
-
-  for (let i = 0; i < txs1.length; i++) {
-    const dateMatch = txs1[i].date === txs2[i].date;
-    const amountMatch = Math.abs(txs1[i].amount - txs2[i].amount) < 0.01;
-
-    if (!dateMatch || !amountMatch) {
-      console.log(`    [Page ${pageNum}] Tx ${i + 1} differs:`);
-      console.log(`      Q1: ${txs1[i].date} | ${txs1[i].amount}`);
-      console.log(`      Q2: ${txs2[i].date} | ${txs2[i].amount}`);
+function extractJsonFromResponse(response: string): unknown[] | null {
+  // Try to find JSON in task_complete tags
+  const completeMatch = response.match(/<task_complete>([\s\S]*?)<\/task_complete>/);
+  if (completeMatch) {
+    const content = completeMatch[1].trim();
+    // Try to find JSON in the content
+    const codeBlockMatch = content.match(/```(?:json)?\s*([\s\S]*?)```/);
+    const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : content;
+    try {
+      const parsed = JSON.parse(jsonStr);
+      if (Array.isArray(parsed)) return parsed;
+    } catch {
+      // Try to find JSON array pattern
+      const jsonMatch = jsonStr.match(/\[[\s\S]*\]/);
+      if (jsonMatch) {
+        try {
+          const parsed = JSON.parse(jsonMatch[0]);
+          if (Array.isArray(parsed)) return parsed;
+        } catch {
+          return null;
+        }
+      }
     }
   }
+
+  // Try to find JSON in markdown code block
+  const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
+  const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
+
+  try {
+    const parsed = JSON.parse(jsonStr);
+    if (Array.isArray(parsed)) return parsed;
+  } catch {
+    // Try to find JSON array pattern
+    const jsonMatch = jsonStr.match(/\[[\s\S]*\]/);
+    if (jsonMatch) {
+      try {
+        const parsed = JSON.parse(jsonMatch[0]);
+        if (Array.isArray(parsed)) return parsed;
+      } catch {
+        return null;
+      }
+    }
+  }
+  return null;
 }
 
 /**
- * Extract transactions from a single page with consensus
+ * Parse JSON response into transactions
+ */
+function parseJsonToTransactions(response: string): ITransaction[] {
+  const parsed = extractJsonFromResponse(response);
+  if (!parsed || !Array.isArray(parsed)) return [];
+
+  return parsed.map((tx: any) => ({
+    date: String(tx.date || ''),
+    counterparty: String(tx.counterparty || tx.description || ''),
+    amount: parseAmount(tx.amount),
+  }));
+}
+
+/**
+ * Extract transactions from a single page using smartagent orchestrator
  */
 async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
-  const MAX_ATTEMPTS = 5;
   console.log(`\n    ======== Page ${pageNum} ========`);
-  console.log(`    [Page ${pageNum}] Starting JSON extraction...`);
 
-  for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt++) {
-    console.log(`\n    [Page ${pageNum}] --- Attempt ${attempt}/${MAX_ATTEMPTS} ---`);
+  const startTime = Date.now();
 
-    // Extract twice in parallel
-    const q1Id = `P${pageNum}A${attempt}Q1`;
-    const q2Id = `P${pageNum}A${attempt}Q2`;
+  const result = await orchestrator.run(EXTRACTION_PROMPT, { images: [image] });
 
-    const [response1, response2] = await Promise.all([
-      queryJson(image, q1Id),
-      queryJson(image, q2Id),
-    ]);
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  console.log(`\n    [Page ${pageNum}] Completed in ${elapsed}s (${result.iterations} iterations, status: ${result.status})`);
 
-    const txs1 = parseJsonResponse(response1, q1Id);
-    const txs2 = parseJsonResponse(response2, q2Id);
+  const transactions = parseJsonToTransactions(result.result);
 
-    console.log(`    [Page ${pageNum}] Results: Q1=${txs1.length} txs, Q2=${txs2.length} txs`);
-
-    if (txs1.length > 0 && transactionArraysMatch(txs1, txs2)) {
-      console.log(`    [Page ${pageNum}] ✓ CONSENSUS REACHED: ${txs1.length} transactions`);
-      console.log(`    [Page ${pageNum}] Transactions:`);
-      for (let i = 0; i < txs1.length; i++) {
-        const tx = txs1[i];
-        console.log(`      ${(i + 1).toString().padStart(2)}. ${tx.date} | ${tx.counterparty.substring(0, 30).padEnd(30)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
-      }
-      return txs1;
-    }
-
-    console.log(`    [Page ${pageNum}] ✗ NO CONSENSUS`);
-    compareAndLogDifferences(txs1, txs2, pageNum);
-
-    if (attempt < MAX_ATTEMPTS) {
-      console.log(`    [Page ${pageNum}] Retrying...`);
-    }
-  }
-
-  // Fallback: use last response
-  console.log(`\n    [Page ${pageNum}] === FALLBACK (no consensus after ${MAX_ATTEMPTS} attempts) ===`);
-  const fallbackId = `P${pageNum}FALLBACK`;
-  const fallbackResponse = await queryJson(image, fallbackId);
-  const fallback = parseJsonResponse(fallbackResponse, fallbackId);
-  console.log(`    [Page ${pageNum}] ~ FALLBACK RESULT: ${fallback.length} transactions`);
-  for (let i = 0; i < fallback.length; i++) {
-    const tx = fallback[i];
+  console.log(`    [Page ${pageNum}] Extracted ${transactions.length} transactions:`);
+  for (let i = 0; i < Math.min(transactions.length, 10); i++) {
+    const tx = transactions[i];
     console.log(`      ${(i + 1).toString().padStart(2)}. ${tx.date} | ${tx.counterparty.substring(0, 30).padEnd(30)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
   }
-  return fallback;
+  if (transactions.length > 10) {
+    console.log(`      ... and ${transactions.length - 10} more transactions`);
+  }
+
+  return transactions;
 }
 
 /**
  * Extract all transactions from bank statement
  */
 async function extractTransactions(images: string[]): Promise<ITransaction[]> {
-  console.log(`    [Vision] Processing ${images.length} page(s) with ${MODEL} (JSON consensus)`);
+  console.log(`    [Vision] Processing ${images.length} page(s) with smartagent DualAgentOrchestrator`);
 
   const allTransactions: ITransaction[] = [];
 
@@ -474,6 +289,80 @@ tap.test('setup: ensure Docker containers are running', async () => {
   console.log('\n[Setup] All containers ready!\n');
 });
 
+tap.test('setup: initialize smartagent orchestrator', async () => {
+  console.log('[Setup] Initializing SmartAi and DualAgentOrchestrator...');
+
+  smartAi = new SmartAi({
+    ollama: {
+      baseUrl: OLLAMA_URL,
+      model: MODEL,
+      defaultOptions: {
+        num_ctx: 32768,
+        num_predict: 4000,
+        temperature: 0.1,
+      },
+      defaultTimeout: 300000, // 5 minutes for vision tasks
+    },
+  });
+
+  await smartAi.start();
+
+  orchestrator = new DualAgentOrchestrator({
+    smartAiInstance: smartAi,
+    defaultProvider: 'ollama',
+    guardianPolicyPrompt: `You are a Guardian agent overseeing bank statement extraction tasks.
+
+APPROVE all tool calls that:
+- Use the json.validate action to verify JSON output
+- Are reasonable attempts to complete the extraction task
+
+REJECT tool calls that:
+- Attempt to access external resources
+- Try to execute arbitrary code
+- Are clearly unrelated to bank statement extraction`,
+    driverSystemMessage: `You are an AI assistant that extracts bank transactions from statement images.
+
+Your task is to analyze bank statement images and extract transaction data.
+You have access to a json.validate tool to verify your JSON output.
+
+IMPORTANT: Always validate your JSON before completing the task.
+
+## Tool Usage Format
+When you need to validate JSON, output:
+
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON_ARRAY"}</params>
+</tool_call>
+
+## Completion Format
+After validation passes, complete the task:
+
+<task_complete>
+[{"date": "YYYY-MM-DD", "counterparty": "...", "amount": -123.45}, ...]
+</task_complete>`,
+    maxIterations: 5,
+    maxConsecutiveRejections: 3,
+    onToken: (token, source) => {
+      if (source === 'driver') {
+        process.stdout.write(token);
+      }
+    },
+    onProgress: (event) => {
+      if (event.logLevel === 'error') {
+        console.error(event.logMessage);
+      }
+    },
+  });
+
+  // Register the JsonValidatorTool
+  orchestrator.registerTool(new JsonValidatorTool());
+
+  await orchestrator.start();
+  console.log('[Setup] Orchestrator initialized!\n');
+});
+
 tap.test('should have MiniCPM-V model loaded', async () => {
   const response = await fetch(`${OLLAMA_URL}/api/tags`);
   const data = await response.json();
@@ -482,7 +371,7 @@ tap.test('should have MiniCPM-V model loaded', async () => {
 });
 
 const testCases = findTestCases();
-console.log(`\nFound ${testCases.length} bank statement test cases (MiniCPM-V)\n`);
+console.log(`\nFound ${testCases.length} bank statement test cases (smartagent + MiniCPM-V)\n`);
 
 let passedCount = 0;
 let failedCount = 0;
@@ -514,7 +403,10 @@ for (const testCase of testCases) {
     // Log counterparty variations (names that differ but date/amount matched)
     if (result.variations.length > 0) {
       console.log(`  Counterparty variations (${result.variations.length}):`);
-      result.variations.forEach((v) => console.log(`    ${v}`));
+      result.variations.slice(0, 5).forEach((v) => console.log(`    ${v}`));
+      if (result.variations.length > 5) {
+        console.log(`    ... and ${result.variations.length - 5} more variations`);
+      }
     }
 
     expect(result.matches).toEqual(result.total);
@@ -522,12 +414,20 @@ for (const testCase of testCases) {
   });
 }
 
+tap.test('cleanup: stop orchestrator', async () => {
+  if (orchestrator) {
+    await orchestrator.stop();
+  }
+  console.log('[Cleanup] Orchestrator stopped');
+});
+
 tap.test('summary', async () => {
   const total = testCases.length;
   console.log(`\n======================================================`);
-  console.log(`   Bank Statement Summary (${MODEL})`);
+  console.log(`   Bank Statement Summary`);
+  console.log(`   (smartagent + ${MODEL})`);
   console.log(`======================================================`);
-  console.log(`  Method:   JSON per-page + consensus`);
+  console.log(`  Method:   DualAgentOrchestrator with vision`);
   console.log(`  Passed:   ${passedCount}/${total}`);
   console.log(`  Failed:   ${failedCount}/${total}`);
   console.log(`======================================================\n`);

test/test.bankstatements.nanonets.ts

@@ -0,0 +1,752 @@
+/**
+ * Bank statement extraction using Nanonets-OCR2-3B + GPT-OSS 20B (sequential two-stage pipeline)
+ *
+ * Stage 1: Nanonets-OCR2-3B converts ALL document pages to markdown (stop after completion)
+ * Stage 2: GPT-OSS 20B extracts structured JSON from saved markdown (after Nanonets stops)
+ *
+ * This approach avoids GPU contention by running services sequentially.
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { execSync } from 'child_process';
+import * as os from 'os';
+import { ensureNanonetsOcr, ensureMiniCpm, isContainerRunning } from './helpers/docker.js';
+import { SmartAi } from '@push.rocks/smartai';
+import { DualAgentOrchestrator, JsonValidatorTool } from '@push.rocks/smartagent';
+
+const NANONETS_URL = 'http://localhost:8000/v1';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR2-3B';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const EXTRACTION_MODEL = 'gpt-oss:20b';
+
+// Temp directory for storing markdown between stages
+const TEMP_MD_DIR = path.join(os.tmpdir(), 'nanonets-markdown');
+
+// SmartAi instance for Ollama with optimized settings
+const smartAi = new SmartAi({
+  ollama: {
+    baseUrl: OLLAMA_URL,
+    model: EXTRACTION_MODEL,
+    defaultOptions: {
+      num_ctx: 32768,   // Larger context for long statements + thinking
+      temperature: 0,   // Deterministic for JSON extraction
+    },
+    defaultTimeout: 600000,  // 10 minute timeout for large documents
+  },
+});
+
+// DualAgentOrchestrator for structured task execution
+let orchestrator: DualAgentOrchestrator;
+
+interface ITransaction {
+  date: string;
+  counterparty: string;
+  amount: number;
+}
+
+interface IImageData {
+  base64: string;
+  width: number;
+  height: number;
+  pageNum: number;
+}
+
+interface ITestCase {
+  name: string;
+  pdfPath: string;
+  jsonPath: string;
+  markdownPath?: string;
+  images?: IImageData[];
+}
+
+// Nanonets-specific prompt for document OCR to markdown
+const NANONETS_OCR_PROMPT = `Extract the text from the above document as if you were reading it naturally.
+Return the tables in html format.
+Return the equations in LaTeX representation.
+If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
+Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
+Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.`;
+
+// JSON extraction prompt for GPT-OSS 20B (sent AFTER the statement text is provided)
+const JSON_EXTRACTION_PROMPT = `Extract ALL transactions from the bank statement. Return ONLY valid JSON array.
+
+WHERE TO FIND DATA:
+- Transactions are typically in TABLES with columns: Date, Description/Counterparty, Debit, Credit, Balance
+- Look for rows with actual money movements, NOT header rows or summary totals
+
+RULES:
+1. date: Convert to YYYY-MM-DD format
+2. counterparty: The name/description of who the money went to/from
+3. amount: NEGATIVE for debits/withdrawals, POSITIVE for credits/deposits
+4. Only include actual transactions, NOT opening/closing balances
+
+JSON array only:
+[{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-25.99}]`;
+
+// Constants for smart batching
+const MAX_VISUAL_TOKENS = 28000;  // ~32K context minus prompt/output headroom
+const PATCH_SIZE = 14;  // Qwen2.5-VL uses 14x14 patches
+
+/**
+ * Estimate visual tokens for an image based on dimensions
+ */
+function estimateVisualTokens(width: number, height: number): number {
+  return Math.ceil((width * height) / (PATCH_SIZE * PATCH_SIZE));
+}
+
+/**
+ * Process images one page at a time for reliability
+ */
+function batchImages(images: IImageData[]): IImageData[][] {
+  // One page per batch for reliable processing
+  return images.map(img => [img]);
+}
+
+/**
+ * Convert PDF to JPEG images using ImageMagick with dimension tracking
+ */
+function convertPdfToImages(pdfPath: string): IImageData[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
+  const outputPattern = path.join(tempDir, 'page-%d.jpg');
+
+  try {
+    execSync(
+      `convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
+      { stdio: 'pipe' }
+    );
+
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
+    const images: IImageData[] = [];
+
+    for (let i = 0; i < files.length; i++) {
+      const file = files[i];
+      const imagePath = path.join(tempDir, file);
+      const imageData = fs.readFileSync(imagePath);
+
+      // Get image dimensions using identify command
+      const dimensions = execSync(`identify -format "%w %h" "${imagePath}"`, { encoding: 'utf-8' }).trim();
+      const [width, height] = dimensions.split(' ').map(Number);
+
+      images.push({
+        base64: imageData.toString('base64'),
+        width,
+        height,
+        pageNum: i + 1,
+      });
+    }
+
+    return images;
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Convert a batch of pages to markdown using Nanonets-OCR-s
+ */
+async function convertBatchToMarkdown(batch: IImageData[]): Promise<string> {
+  const startTime = Date.now();
+  const pageNums = batch.map(img => img.pageNum).join(', ');
+
+  // Build content array with all images first, then the prompt
+  const content: Array<{ type: string; image_url?: { url: string }; text?: string }> = [];
+
+  for (const img of batch) {
+    content.push({
+      type: 'image_url',
+      image_url: { url: `data:image/jpeg;base64,${img.base64}` },
+    });
+  }
+
+  // Add prompt with page separator instruction if multiple pages
+  const promptText = batch.length > 1
+    ? `${NANONETS_OCR_PROMPT}\n\nPlease clearly separate each page's content with "--- PAGE N ---" markers, where N is the page number starting from ${batch[0].pageNum}.`
+    : NANONETS_OCR_PROMPT;
+
+  content.push({ type: 'text', text: promptText });
+
+  const response = await fetch(`${NANONETS_URL}/chat/completions`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': 'Bearer dummy',
+    },
+    body: JSON.stringify({
+      model: NANONETS_MODEL,
+      messages: [{
+        role: 'user',
+        content,
+      }],
+      max_tokens: 4096 * batch.length,  // Scale output tokens with batch size
+      temperature: 0.0,
+    }),
+    signal: AbortSignal.timeout(600000),  // 10 minute timeout for OCR
+  });
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+
+  if (!response.ok) {
+    const errorText = await response.text();
+    throw new Error(`Nanonets API error: ${response.status} - ${errorText}`);
+  }
+
+  const data = await response.json();
+  let responseContent = (data.choices?.[0]?.message?.content || '').trim();
+
+  // For single-page batches, add page marker if not present
+  if (batch.length === 1 && !responseContent.includes('--- PAGE')) {
+    responseContent = `--- PAGE ${batch[0].pageNum} ---\n${responseContent}`;
+  }
+
+  console.log(`        Pages [${pageNums}]: ${responseContent.length} chars (${elapsed}s)`);
+  return responseContent;
+}
+
+/**
+ * Convert all pages of a document to markdown using smart batching
+ */
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
+  const batches = batchImages(images);
+  console.log(`      [${docName}] Processing ${images.length} page(s) in ${batches.length} batch(es)...`);
+
+  const markdownParts: string[] = [];
+
+  for (let i = 0; i < batches.length; i++) {
+    const batch = batches[i];
+    const batchTokens = batch.reduce((sum, img) => sum + estimateVisualTokens(img.width, img.height), 0);
+    console.log(`        Batch ${i + 1}: ${batch.length} page(s), ~${batchTokens} tokens`);
+    const markdown = await convertBatchToMarkdown(batch);
+    markdownParts.push(markdown);
+  }
+
+  const fullMarkdown = markdownParts.join('\n\n');
+  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
+  return fullMarkdown;
+}
+
+/**
+ * Stop Nanonets container
+ */
+function stopNanonets(): void {
+  console.log('    [Docker] Stopping Nanonets container...');
+  try {
+    execSync('docker stop nanonets-test 2>/dev/null || true', { stdio: 'pipe' });
+    // Wait for GPU memory to be released
+    execSync('sleep 5', { stdio: 'pipe' });
+    console.log('    [Docker] Nanonets stopped');
+  } catch {
+    console.log('    [Docker] Nanonets was not running');
+  }
+}
+
+/**
+ * Ensure GPT-OSS 20B model is available and warmed up
+ */
+async function ensureExtractionModel(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_URL}/api/tags`);
+    if (response.ok) {
+      const data = await response.json();
+      const models = data.models || [];
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`    [Ollama] Pulling ${EXTRACTION_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: EXTRACTION_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+/**
+ * Try to extract valid JSON from a response string
+ */
+function tryExtractJson(response: string): unknown[] | null {
+  // Remove thinking tags
+  let clean = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+
+  // Try task_complete tags first
+  const completeMatch = clean.match(/<task_complete>([\s\S]*?)<\/task_complete>/);
+  if (completeMatch) {
+    clean = completeMatch[1].trim();
+  }
+
+  // Try code block
+  const codeBlockMatch = clean.match(/```(?:json)?\s*([\s\S]*?)```/);
+  const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : clean;
+
+  try {
+    const parsed = JSON.parse(jsonStr);
+    if (Array.isArray(parsed)) return parsed;
+  } catch {
+    // Try to find JSON array
+    const jsonMatch = jsonStr.match(/\[[\s\S]*\]/);
+    if (jsonMatch) {
+      try {
+        const parsed = JSON.parse(sanitizeJson(jsonMatch[0]));
+        if (Array.isArray(parsed)) return parsed;
+      } catch {
+        return null;
+      }
+    }
+    return null;
+  }
+  return null;
+}
+
+/**
+ * Extract transactions from markdown using smartagent DualAgentOrchestrator
+ * Validates JSON and retries if invalid
+ */
+async function extractTransactionsFromMarkdown(markdown: string, queryId: string): Promise<ITransaction[]> {
+  const startTime = Date.now();
+
+  console.log(`      [${queryId}] Statement: ${markdown.length} chars`);
+
+  // Build the extraction task with document context
+  const taskPrompt = `Extract all transactions from this bank statement document and output ONLY the JSON array:
+
+${markdown}
+
+${JSON_EXTRACTION_PROMPT}
+
+Before completing, validate your JSON using the json.validate tool:
+
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON_ARRAY_HERE"}</params>
+</tool_call>
+
+Only complete after validation passes. Output the final JSON array in <task_complete></task_complete> tags.`;
+
+  try {
+    const result = await orchestrator.run(taskPrompt);
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] Status: ${result.status}, Iterations: ${result.iterations} (${elapsed}s)`);
+
+    // Try to parse JSON from result
+    let jsonData: unknown[] | null = null;
+    let responseText = result.result || '';
+
+    if (result.success && responseText) {
+      jsonData = tryExtractJson(responseText);
+    }
+
+    // Fallback: try parsing from history
+    if (!jsonData && result.history?.length > 0) {
+      const lastMessage = result.history[result.history.length - 1];
+      if (lastMessage?.content) {
+        responseText = lastMessage.content;
+        jsonData = tryExtractJson(responseText);
+      }
+    }
+
+    if (!jsonData) {
+      console.log(`      [${queryId}] Failed to parse JSON`);
+      return [];
+    }
+
+    // Convert to transactions
+    const txs = jsonData.map((tx: any) => ({
+      date: String(tx.date || ''),
+      counterparty: String(tx.counterparty || tx.description || ''),
+      amount: parseAmount(tx.amount),
+    }));
+    console.log(`      [${queryId}] Parsed ${txs.length} transactions`);
+    return txs;
+  } catch (error) {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] ERROR: ${error} (${elapsed}s)`);
+    throw error;
+  }
+}
+
+/**
+ * Sanitize JSON string
+ */
+function sanitizeJson(jsonStr: string): string {
+  let s = jsonStr;
+  s = s.replace(/"amount"\s*:\s*\+/g, '"amount": ');
+  s = s.replace(/:\s*\+(\d)/g, ': $1');
+  s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3.$4');
+  s = s.replace(/,\s*([}\]])/g, '$1');
+  s = s.replace(/"([^"\\]*)\n([^"]*)"/g, '"$1 $2"');
+  s = s.replace(/"([^"\\]*)\t([^"]*)"/g, '"$1 $2"');
+  s = s.replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, ' ');
+  return s;
+}
+
+/**
+ * Parse amount from various formats
+ */
+function parseAmount(value: unknown): number {
+  if (typeof value === 'number') return value;
+  if (typeof value !== 'string') return 0;
+
+  let s = value.replace(/[€$£\s]/g, '').replace('−', '-').replace('–', '-');
+  if (s.includes(',') && s.indexOf(',') > s.lastIndexOf('.')) {
+    s = s.replace(/\./g, '').replace(',', '.');
+  } else {
+    s = s.replace(/,/g, '');
+  }
+  return parseFloat(s) || 0;
+}
+
+/**
+ * Parse JSON response into transactions
+ */
+function parseJsonResponse(response: string, queryId: string): ITransaction[] {
+  // Remove thinking tags if present
+  let cleanResponse = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+
+  // Debug: show what we're working with
+  console.log(`      [${queryId}] Response preview: ${cleanResponse.substring(0, 300)}...`);
+
+  const codeBlockMatch = cleanResponse.match(/```(?:json)?\s*([\s\S]*?)```/);
+  let jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : cleanResponse;
+  jsonStr = sanitizeJson(jsonStr);
+
+  try {
+    const parsed = JSON.parse(jsonStr);
+    if (Array.isArray(parsed)) {
+      const txs = parsed.map(tx => ({
+        date: String(tx.date || ''),
+        counterparty: String(tx.counterparty || tx.description || ''),
+        amount: parseAmount(tx.amount),
+      }));
+      console.log(`      [${queryId}] Parsed ${txs.length} transactions`);
+      return txs;
+    }
+  } catch (e) {
+    // Try to find a JSON array in the text
+    const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
+    if (arrayMatch) {
+      console.log(`      [${queryId}] Array match found: ${arrayMatch[0].length} chars`);
+      try {
+        const parsed = JSON.parse(sanitizeJson(arrayMatch[0]));
+        if (Array.isArray(parsed)) {
+          const txs = parsed.map(tx => ({
+            date: String(tx.date || ''),
+            counterparty: String(tx.counterparty || tx.description || ''),
+            amount: parseAmount(tx.amount),
+          }));
+          console.log(`      [${queryId}] Parsed ${txs.length} transactions (array match)`);
+          return txs;
+        }
+      } catch (innerErr) {
+        console.log(`      [${queryId}] Array parse error: ${(innerErr as Error).message}`);
+      }
+    } else {
+      console.log(`      [${queryId}] No JSON array found in response`);
+    }
+  }
+
+  console.log(`      [${queryId}] PARSE FAILED`);
+  return [];
+}
+
+/**
+ * Extract transactions (single pass)
+ */
+async function extractTransactions(markdown: string, docName: string): Promise<ITransaction[]> {
+  console.log(`    [${docName}] Extracting...`);
+  const txs = await extractTransactionsFromMarkdown(markdown, docName);
+  console.log(`    [${docName}] Extracted ${txs.length} transactions`);
+  return txs;
+}
+
+/**
+ * Compare transactions
+ */
+function compareTransactions(
+  extracted: ITransaction[],
+  expected: ITransaction[]
+): { matches: number; total: number; errors: string[] } {
+  const errors: string[] = [];
+  let matches = 0;
+
+  for (let i = 0; i < expected.length; i++) {
+    const exp = expected[i];
+    const ext = extracted[i];
+
+    if (!ext) {
+      errors.push(`Missing tx ${i}: ${exp.date} ${exp.counterparty}`);
+      continue;
+    }
+
+    const dateMatch = ext.date === exp.date;
+    const amountMatch = Math.abs(ext.amount - exp.amount) < 0.01;
+
+    if (dateMatch && amountMatch) {
+      matches++;
+    } else {
+      errors.push(`Mismatch ${i}: exp ${exp.date}/${exp.amount}, got ${ext.date}/${ext.amount}`);
+    }
+  }
+
+  if (extracted.length > expected.length) {
+    errors.push(`Extra transactions: ${extracted.length - expected.length}`);
+  }
+
+  return { matches, total: expected.length, errors };
+}
+
+/**
+ * Find all test cases
+ */
+function findTestCases(): ITestCase[] {
+  const testDir = path.join(process.cwd(), '.nogit');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: ITestCase[] = [];
+
+  for (const pdf of files.filter((f: string) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${baseName}.json`;
+    if (files.includes(jsonFile)) {
+      testCases.push({
+        name: baseName,
+        pdfPath: path.join(testDir, pdf),
+        jsonPath: path.join(testDir, jsonFile),
+      });
+    }
+  }
+
+  return testCases.sort((a, b) => a.name.localeCompare(b.name));
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} bank statement test cases\n`);
+
+// Ensure temp directory exists
+if (!fs.existsSync(TEMP_MD_DIR)) {
+  fs.mkdirSync(TEMP_MD_DIR, { recursive: true });
+}
+
+// -------- STAGE 1: OCR with Nanonets --------
+
+// Check if all markdown files already exist
+function allMarkdownFilesExist(): boolean {
+  for (const tc of testCases) {
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Track whether we need to run Stage 1
+let stage1Needed = !allMarkdownFilesExist();
+
+tap.test('Stage 1: Setup Nanonets', async () => {
+  console.log('\n========== STAGE 1: Nanonets OCR ==========\n');
+
+  if (!stage1Needed) {
+    console.log('  [SKIP] All markdown files already exist, skipping Nanonets setup');
+    return;
+  }
+
+  const ok = await ensureNanonetsOcr();
+  expect(ok).toBeTrue();
+});
+
+tap.test('Stage 1: Convert all documents to markdown', async () => {
+  if (!stage1Needed) {
+    console.log('  [SKIP] Using existing markdown files from previous run\n');
+    // Load existing markdown paths
+    for (const tc of testCases) {
+      tc.markdownPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+      console.log(`    Loaded: ${tc.markdownPath}`);
+    }
+    return;
+  }
+
+  console.log('\n  Converting all PDFs to markdown with Nanonets-OCR-s...\n');
+
+  for (const tc of testCases) {
+    console.log(`\n    === ${tc.name} ===`);
+
+    // Convert PDF to images
+    const images = convertPdfToImages(tc.pdfPath);
+    console.log(`      Pages: ${images.length}`);
+
+    // Convert to markdown
+    const markdown = await convertDocumentToMarkdown(images, tc.name);
+
+    // Save markdown to temp file
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    fs.writeFileSync(mdPath, markdown);
+    tc.markdownPath = mdPath;
+    console.log(`      Saved: ${mdPath}`);
+  }
+
+  console.log('\n  Stage 1 complete: All documents converted to markdown\n');
+});
+
+tap.test('Stage 1: Stop Nanonets', async () => {
+  if (!stage1Needed) {
+    console.log('  [SKIP] Nanonets was not started');
+    return;
+  }
+
+  stopNanonets();
+  // Verify it's stopped
+  await new Promise(resolve => setTimeout(resolve, 3000));
+  expect(isContainerRunning('nanonets-test')).toBeFalse();
+});
+
+// -------- STAGE 2: Extraction with GPT-OSS 20B --------
+
+tap.test('Stage 2: Setup Ollama + GPT-OSS 20B', async () => {
+  console.log('\n========== STAGE 2: GPT-OSS 20B Extraction ==========\n');
+
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
+
+  // Initialize SmartAi and DualAgentOrchestrator
+  console.log('  [SmartAgent] Starting SmartAi...');
+  await smartAi.start();
+
+  console.log('  [SmartAgent] Creating DualAgentOrchestrator...');
+  orchestrator = new DualAgentOrchestrator({
+    smartAiInstance: smartAi,
+    defaultProvider: 'ollama',
+    guardianPolicyPrompt: `
+      JSON EXTRACTION POLICY:
+      - APPROVE all JSON extraction tasks
+      - APPROVE all json.validate tool calls
+      - This is a read-only operation - no file system or network access needed
+      - The task is to extract structured transaction data from document text
+    `,
+    driverSystemMessage: `You are a precise JSON extraction assistant. Your only job is to extract transaction data from bank statements.
+
+CRITICAL RULES:
+1. Output valid JSON array with the exact format requested
+2. Amounts should be NEGATIVE for debits/withdrawals, POSITIVE for credits/deposits
+3. IMPORTANT: Before completing, validate your JSON using the json.validate tool:
+
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON_ARRAY"}</params>
+</tool_call>
+
+4. Only complete after validation passes
+
+When done, wrap your JSON array in <task_complete></task_complete> tags.`,
+    maxIterations: 5,
+    // Enable streaming for real-time progress visibility
+    onToken: (token, source) => {
+      if (source === 'driver') {
+        process.stdout.write(token);
+      }
+    },
+  });
+
+  // Register JsonValidatorTool for self-validation
+  orchestrator.registerTool(new JsonValidatorTool());
+
+  console.log('  [SmartAgent] Starting orchestrator...');
+  await orchestrator.start();
+  console.log('  [SmartAgent] Ready for extraction');
+});
+
+let passedCount = 0;
+let failedCount = 0;
+
+for (const tc of testCases) {
+  tap.test(`Stage 2: Extract ${tc.name}`, async () => {
+    const expected: ITransaction[] = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    console.log(`\n    === ${tc.name} ===`);
+    console.log(`    Expected: ${expected.length} transactions`);
+
+    // Load saved markdown
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      throw new Error(`Markdown not found: ${mdPath}. Run Stage 1 first.`);
+    }
+    const markdown = fs.readFileSync(mdPath, 'utf-8');
+    console.log(`    Markdown: ${markdown.length} chars`);
+
+    // Extract transactions (single pass)
+    const extracted = await extractTransactions(markdown, tc.name);
+
+    // Log results
+    console.log(`    Extracted: ${extracted.length} transactions`);
+    for (let i = 0; i < Math.min(extracted.length, 5); i++) {
+      const tx = extracted[i];
+      console.log(`      ${i + 1}. ${tx.date} | ${tx.counterparty.substring(0, 25).padEnd(25)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
+    }
+    if (extracted.length > 5) {
+      console.log(`      ... and ${extracted.length - 5} more`);
+    }
+
+    // Compare
+    const result = compareTransactions(extracted, expected);
+    const pass = result.matches === result.total && extracted.length === expected.length;
+
+    if (pass) {
+      passedCount++;
+      console.log(`    Result: PASS (${result.matches}/${result.total})`);
+    } else {
+      failedCount++;
+      console.log(`    Result: FAIL (${result.matches}/${result.total})`);
+      result.errors.slice(0, 5).forEach(e => console.log(`      - ${e}`));
+    }
+
+    expect(result.matches).toEqual(result.total);
+    expect(extracted.length).toEqual(expected.length);
+  });
+}
+
+tap.test('Summary', async () => {
+  // Cleanup orchestrator and SmartAi
+  if (orchestrator) {
+    console.log('\n  [SmartAgent] Stopping orchestrator...');
+    await orchestrator.stop();
+  }
+  console.log('  [SmartAgent] Stopping SmartAi...');
+  await smartAi.stop();
+
+  console.log(`\n======================================================`);
+  console.log(`   Bank Statement Summary (Nanonets + SmartAgent)`);
+  console.log(`======================================================`);
+  console.log(`  Stage 1:  Nanonets-OCR-s (document -> markdown)`);
+  console.log(`  Stage 2:  GPT-OSS 20B + SmartAgent (markdown -> JSON)`);
+  console.log(`  Passed:   ${passedCount}/${testCases.length}`);
+  console.log(`  Failed:   ${failedCount}/${testCases.length}`);
+  console.log(`======================================================\n`);
+
+  // Only cleanup temp files if ALL tests passed
+  if (failedCount === 0 && passedCount === testCases.length) {
+    try {
+      fs.rmSync(TEMP_MD_DIR, { recursive: true, force: true });
+      console.log(`  Cleaned up temp directory: ${TEMP_MD_DIR}\n`);
+    } catch {
+      // Ignore
+    }
+  } else {
+    console.log(`  Keeping temp directory for debugging: ${TEMP_MD_DIR}\n`);
+  }
+});
+
+export default tap.start();

test/test.invoices.extraction.ts

@@ -0,0 +1,440 @@
+/**
+ * Invoice extraction tuning - uses pre-generated markdown files
+ *
+ * Skips OCR stage, only runs GPT-OSS extraction on existing .debug.md files.
+ * Use this to quickly iterate on extraction prompts and logic.
+ *
+ * Run with: tstest test/test.invoices.extraction.ts --verbose
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { ensureMiniCpm } from './helpers/docker.js';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const EXTRACTION_MODEL = 'gpt-oss:20b';
+
+// Test these specific invoices (must have .debug.md files)
+const TEST_INVOICES = [
+  'consensus_2021-09',
+  'hetzner_2022-04',
+  'qonto_2021-08',
+  'qonto_2021-09',
+];
+
+interface IInvoice {
+  invoice_number: string;
+  invoice_date: string;
+  vendor_name: string;
+  currency: string;
+  net_amount: number;
+  vat_amount: number;
+  total_amount: number;
+}
+
+interface ITestCase {
+  name: string;
+  markdownPath: string;
+  jsonPath: string;
+}
+
+// JSON extraction prompt for GPT-OSS 20B (sent AFTER the invoice text is provided)
+const JSON_EXTRACTION_PROMPT = `Extract key fields from the invoice. Return ONLY valid JSON.
+
+WHERE TO FIND DATA:
+- invoice_number, invoice_date, vendor_name: Look in the HEADER section at the TOP of PAGE 1 (near "Invoice no.", "Invoice date:", "Rechnungsnummer")
+- net_amount, vat_amount, total_amount: Look in the SUMMARY section at the BOTTOM (look for "Total", "Amount due", "Gesamtbetrag")
+
+RULES:
+1. invoice_number: Extract ONLY the value (e.g., "R0015632540"), NOT the label "Invoice no."
+2. invoice_date: Convert to YYYY-MM-DD format (e.g., "14/04/2022" → "2022-04-14")
+3. vendor_name: The company issuing the invoice
+4. currency: EUR, USD, or GBP
+5. net_amount: Total before tax
+6. vat_amount: Tax amount
+7. total_amount: Final total with tax
+
+JSON only:
+{"invoice_number":"X","invoice_date":"YYYY-MM-DD","vendor_name":"X","currency":"EUR","net_amount":0,"vat_amount":0,"total_amount":0}`;
+
+/**
+ * Ensure GPT-OSS 20B model is available
+ */
+async function ensureExtractionModel(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_URL}/api/tags`);
+    if (response.ok) {
+      const data = await response.json();
+      const models = data.models || [];
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`    [Ollama] Pulling ${EXTRACTION_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: EXTRACTION_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+/**
+ * Parse amount from string (handles European format)
+ */
+function parseAmount(s: string | number | undefined): number {
+  if (s === undefined || s === null) return 0;
+  if (typeof s === 'number') return s;
+  const match = s.match(/([\d.,]+)/);
+  if (!match) return 0;
+  const numStr = match[1];
+  const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
+    ? numStr.replace(/\./g, '').replace(',', '.')
+    : numStr.replace(/,/g, '');
+  return parseFloat(normalized) || 0;
+}
+
+/**
+ * Extract invoice number - minimal normalization
+ */
+function extractInvoiceNumber(s: string | undefined): string {
+  if (!s) return '';
+  return s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+}
+
+/**
+ * Extract date (YYYY-MM-DD) from response
+ */
+function extractDate(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
+  if (isoMatch) return isoMatch[1];
+  const dmyMatch = clean.match(/(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})/);
+  if (dmyMatch) {
+    return `${dmyMatch[3]}-${dmyMatch[2].padStart(2, '0')}-${dmyMatch[1].padStart(2, '0')}`;
+  }
+  return clean.replace(/[^\d-]/g, '').trim();
+}
+
+/**
+ * Extract currency
+ */
+function extractCurrency(s: string | undefined): string {
+  if (!s) return 'EUR';
+  const upper = s.toUpperCase();
+  if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
+  if (upper.includes('USD') || upper.includes('$')) return 'USD';
+  if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
+  return 'EUR';
+}
+
+/**
+ * Extract JSON from response
+ */
+function extractJsonFromResponse(response: string): Record<string, unknown> | null {
+  let cleanResponse = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+  const codeBlockMatch = cleanResponse.match(/```(?:json)?\s*([\s\S]*?)```/);
+  const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : cleanResponse;
+
+  try {
+    return JSON.parse(jsonStr);
+  } catch {
+    const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
+    if (jsonMatch) {
+      try {
+        return JSON.parse(jsonMatch[0]);
+      } catch {
+        return null;
+      }
+    }
+    return null;
+  }
+}
+
+/**
+ * Parse JSON response into IInvoice
+ */
+function parseJsonToInvoice(response: string): IInvoice | null {
+  const parsed = extractJsonFromResponse(response);
+  if (!parsed) return null;
+
+  return {
+    invoice_number: extractInvoiceNumber(String(parsed.invoice_number || '')),
+    invoice_date: extractDate(String(parsed.invoice_date || '')),
+    vendor_name: String(parsed.vendor_name || '').replace(/\*\*/g, '').replace(/`/g, '').trim(),
+    currency: extractCurrency(String(parsed.currency || '')),
+    net_amount: parseAmount(parsed.net_amount as string | number),
+    vat_amount: parseAmount(parsed.vat_amount as string | number),
+    total_amount: parseAmount(parsed.total_amount as string | number),
+  };
+}
+
+/**
+ * Extract invoice from markdown using GPT-OSS 20B (streaming)
+ */
+async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Promise<IInvoice | null> {
+  const startTime = Date.now();
+
+  console.log(`      [${queryId}] Invoice: ${markdown.length} chars, Prompt: ${JSON_EXTRACTION_PROMPT.length} chars`);
+
+  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: EXTRACTION_MODEL,
+      messages: [
+        { role: 'user', content: 'Hi there, how are you?' },
+        { role: 'assistant', content: 'Good, how can I help you today?' },
+        { role: 'user', content: `Here is an invoice document:\n\n${markdown}` },
+        { role: 'assistant', content: 'I have read the invoice document you provided. I can see all the text content. What would you like me to do with it?' },
+        { role: 'user', content: JSON_EXTRACTION_PROMPT },
+      ],
+      stream: true,
+      options: {
+        num_ctx: 32768,  // Larger context for long invoices + thinking
+        temperature: 0,  // Deterministic for JSON extraction
+      },
+    }),
+    signal: AbortSignal.timeout(120000), // 2 min timeout
+  });
+
+  if (!response.ok) {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] ERROR: ${response.status} (${elapsed}s)`);
+    throw new Error(`Ollama API error: ${response.status}`);
+  }
+
+  // Stream the response
+  let content = '';
+  let thinkingContent = '';
+  let thinkingStarted = false;
+  let outputStarted = false;
+  const reader = response.body!.getReader();
+  const decoder = new TextDecoder();
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+
+      const chunk = decoder.decode(value, { stream: true });
+
+      for (const line of chunk.split('\n').filter(l => l.trim())) {
+        try {
+          const json = JSON.parse(line);
+
+          const thinking = json.message?.thinking || '';
+          if (thinking) {
+            if (!thinkingStarted) {
+              process.stdout.write(`      [${queryId}] THINKING: `);
+              thinkingStarted = true;
+            }
+            process.stdout.write(thinking);
+            thinkingContent += thinking;
+          }
+
+          const token = json.message?.content || '';
+          if (token) {
+            if (!outputStarted) {
+              if (thinkingStarted) process.stdout.write('\n');
+              process.stdout.write(`      [${queryId}] OUTPUT: `);
+              outputStarted = true;
+            }
+            process.stdout.write(token);
+            content += token;
+          }
+        } catch {
+          // Ignore parse errors for partial chunks
+        }
+      }
+    }
+  } finally {
+    if (thinkingStarted || outputStarted) process.stdout.write('\n');
+  }
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  console.log(`      [${queryId}] Done: ${thinkingContent.length} thinking, ${content.length} output (${elapsed}s)`);
+
+  return parseJsonToInvoice(content);
+}
+
+/**
+ * Normalize date to YYYY-MM-DD
+ */
+function normalizeDate(dateStr: string | null): string {
+  if (!dateStr) return '';
+  if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
+
+  const monthMap: Record<string, string> = {
+    JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
+    JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
+  };
+
+  let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
+  if (match) {
+    return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
+  }
+
+  match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
+  if (match) {
+    return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
+  }
+
+  return dateStr;
+}
+
+/**
+ * Normalize invoice number for comparison (remove spaces, lowercase)
+ */
+function normalizeInvoiceNumber(s: string): string {
+  return s.replace(/\s+/g, '').toLowerCase();
+}
+
+/**
+ * Compare extracted invoice against expected
+ */
+function compareInvoice(
+  extracted: IInvoice,
+  expected: IInvoice
+): { match: boolean; errors: string[] } {
+  const errors: string[] = [];
+
+  // Invoice number - normalize spaces for comparison
+  const extNum = normalizeInvoiceNumber(extracted.invoice_number || '');
+  const expNum = normalizeInvoiceNumber(expected.invoice_number || '');
+  if (extNum !== expNum) {
+    errors.push(`invoice_number: expected "${expected.invoice_number}", got "${extracted.invoice_number}"`);
+  }
+
+  if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
+    errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
+  }
+
+  if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
+    errors.push(`total_amount: expected ${expected.total_amount}, got ${extracted.total_amount}`);
+  }
+
+  if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
+    errors.push(`currency: expected "${expected.currency}", got "${extracted.currency}"`);
+  }
+
+  return { match: errors.length === 0, errors };
+}
+
+/**
+ * Find test cases with existing debug markdown
+ */
+function findTestCases(): ITestCase[] {
+  const invoicesDir = path.join(process.cwd(), '.nogit/invoices');
+  if (!fs.existsSync(invoicesDir)) return [];
+
+  const testCases: ITestCase[] = [];
+
+  for (const invoiceName of TEST_INVOICES) {
+    const markdownPath = path.join(invoicesDir, `${invoiceName}.debug.md`);
+    const jsonPath = path.join(invoicesDir, `${invoiceName}.json`);
+
+    if (fs.existsSync(markdownPath) && fs.existsSync(jsonPath)) {
+      testCases.push({
+        name: invoiceName,
+        markdownPath,
+        jsonPath,
+      });
+    } else {
+      if (!fs.existsSync(markdownPath)) {
+        console.warn(`Warning: Missing markdown: ${markdownPath}`);
+      }
+      if (!fs.existsSync(jsonPath)) {
+        console.warn(`Warning: Missing JSON: ${jsonPath}`);
+      }
+    }
+  }
+
+  return testCases;
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\n========================================`);
+console.log(`  EXTRACTION TUNING TEST`);
+console.log(`  (Skips OCR, uses existing .debug.md)`);
+console.log(`========================================`);
+console.log(`  Testing ${testCases.length} invoices:`);
+for (const tc of testCases) {
+  console.log(`    - ${tc.name}`);
+}
+console.log(`========================================\n`);
+
+tap.test('Setup Ollama + GPT-OSS 20B', async () => {
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
+});
+
+let passedCount = 0;
+let failedCount = 0;
+
+for (const tc of testCases) {
+  tap.test(`Extract ${tc.name}`, async () => {
+    const expected: IInvoice = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    const markdown = fs.readFileSync(tc.markdownPath, 'utf-8');
+
+    console.log(`\n    ========================================`);
+    console.log(`    === ${tc.name} ===`);
+    console.log(`    ========================================`);
+    console.log(`    EXPECTED: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
+    console.log(`    Markdown: ${markdown.length} chars`);
+
+    const startTime = Date.now();
+
+    const extracted = await extractInvoiceFromMarkdown(markdown, tc.name);
+
+    if (!extracted) {
+      failedCount++;
+      console.log(`\n    Result: ✗ FAILED TO PARSE (${((Date.now() - startTime) / 1000).toFixed(1)}s)`);
+      return;
+    }
+
+    const elapsedMs = Date.now() - startTime;
+
+    console.log(`    EXTRACTED: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
+
+    const result = compareInvoice(extracted, expected);
+
+    if (result.match) {
+      passedCount++;
+      console.log(`\n    Result: ✓ MATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+    } else {
+      failedCount++;
+      console.log(`\n    Result: ✗ MISMATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+      console.log(`    ERRORS:`);
+      result.errors.forEach(e => console.log(`      - ${e}`));
+    }
+  });
+}
+
+tap.test('Summary', async () => {
+  const totalInvoices = testCases.length;
+  const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
+
+  console.log(`\n========================================`);
+  console.log(`   Extraction Tuning Summary`);
+  console.log(`========================================`);
+  console.log(`  Model:     ${EXTRACTION_MODEL}`);
+  console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
+  console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
+  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
+  console.log(`========================================\n`);
+});
+
+export default tap.start();

test/test.invoices.failed.ts

@@ -0,0 +1,695 @@
+/**
+ * Focused test for failed invoice extractions
+ *
+ * Tests only the 4 invoices that failed in the main test:
+ * - consensus_2021-09: invoice_number "2021/1384" → "20211384" (slash stripped)
+ * - hetzner_2022-04: model hallucinated after 281s thinking
+ * - qonto_2021-08: invoice_number "08-21-INVOICE-410870" → "4108705" (prefix stripped)
+ * - qonto_2021-09: invoice_number "09-21-INVOICE-4303642" → "4303642" (prefix stripped)
+ *
+ * Run with: tstest test/test.invoices.failed.ts --verbose
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { execSync } from 'child_process';
+import * as os from 'os';
+import { ensureNanonetsOcr, ensureMiniCpm, isContainerRunning } from './helpers/docker.js';
+
+const NANONETS_URL = 'http://localhost:8000/v1';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR2-3B';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const EXTRACTION_MODEL = 'gpt-oss:20b';
+
+// Temp directory for storing markdown between stages
+const TEMP_MD_DIR = path.join(os.tmpdir(), 'nanonets-invoices-failed-debug');
+
+// Only test these specific invoices that failed
+const FAILED_INVOICES = [
+  'consensus_2021-09',
+  'hetzner_2022-04',
+  'qonto_2021-08',
+  'qonto_2021-09',
+];
+
+interface IInvoice {
+  invoice_number: string;
+  invoice_date: string;
+  vendor_name: string;
+  currency: string;
+  net_amount: number;
+  vat_amount: number;
+  total_amount: number;
+}
+
+interface IImageData {
+  base64: string;
+  width: number;
+  height: number;
+  pageNum: number;
+}
+
+interface ITestCase {
+  name: string;
+  pdfPath: string;
+  jsonPath: string;
+  markdownPath?: string;
+}
+
+// Nanonets-specific prompt for document OCR to markdown
+const NANONETS_OCR_PROMPT = `Extract the text from the above document as if you were reading it naturally.
+Return the tables in html format.
+Return the equations in LaTeX representation.
+If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
+Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
+Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.`;
+
+// JSON extraction prompt for GPT-OSS 20B
+const JSON_EXTRACTION_PROMPT = `You are an invoice data extractor. Below is an invoice document converted to text/markdown. Extract the key invoice fields as JSON.
+
+IMPORTANT RULES:
+1. invoice_number: The unique invoice/document number (NOT VAT ID, NOT customer ID). PRESERVE ALL CHARACTERS including slashes, dashes, and prefixes.
+2. invoice_date: Format as YYYY-MM-DD
+3. vendor_name: The company that issued the invoice
+4. currency: EUR, USD, or GBP
+5. net_amount: Amount before tax
+6. vat_amount: Tax/VAT amount
+7. total_amount: Final total (gross amount)
+
+Return ONLY this JSON format, no explanation:
+{
+  "invoice_number": "INV-2024-001",
+  "invoice_date": "2024-01-15",
+  "vendor_name": "Company Name",
+  "currency": "EUR",
+  "net_amount": 100.00,
+  "vat_amount": 19.00,
+  "total_amount": 119.00
+}
+
+INVOICE TEXT:
+`;
+
+const PATCH_SIZE = 14;
+
+/**
+ * Estimate visual tokens for an image based on dimensions
+ */
+function estimateVisualTokens(width: number, height: number): number {
+  return Math.ceil((width * height) / (PATCH_SIZE * PATCH_SIZE));
+}
+
+/**
+ * Process images one page at a time for reliability
+ */
+function batchImages(images: IImageData[]): IImageData[][] {
+  return images.map(img => [img]);
+}
+
+/**
+ * Convert PDF to JPEG images using ImageMagick with dimension tracking
+ */
+function convertPdfToImages(pdfPath: string): IImageData[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
+  const outputPattern = path.join(tempDir, 'page-%d.jpg');
+
+  try {
+    execSync(
+      `convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
+      { stdio: 'pipe' }
+    );
+
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
+    const images: IImageData[] = [];
+
+    for (let i = 0; i < files.length; i++) {
+      const file = files[i];
+      const imagePath = path.join(tempDir, file);
+      const imageData = fs.readFileSync(imagePath);
+
+      const dimensions = execSync(`identify -format "%w %h" "${imagePath}"`, { encoding: 'utf-8' }).trim();
+      const [width, height] = dimensions.split(' ').map(Number);
+
+      images.push({
+        base64: imageData.toString('base64'),
+        width,
+        height,
+        pageNum: i + 1,
+      });
+    }
+
+    return images;
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Convert a batch of pages to markdown using Nanonets-OCR-s
+ */
+async function convertBatchToMarkdown(batch: IImageData[]): Promise<string> {
+  const startTime = Date.now();
+  const pageNums = batch.map(img => img.pageNum).join(', ');
+
+  const content: Array<{ type: string; image_url?: { url: string }; text?: string }> = [];
+
+  for (const img of batch) {
+    content.push({
+      type: 'image_url',
+      image_url: { url: `data:image/jpeg;base64,${img.base64}` },
+    });
+  }
+
+  const promptText = batch.length > 1
+    ? `${NANONETS_OCR_PROMPT}\n\nPlease clearly separate each page's content with "--- PAGE N ---" markers, where N is the page number starting from ${batch[0].pageNum}.`
+    : NANONETS_OCR_PROMPT;
+
+  content.push({ type: 'text', text: promptText });
+
+  const response = await fetch(`${NANONETS_URL}/chat/completions`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': 'Bearer dummy',
+    },
+    body: JSON.stringify({
+      model: NANONETS_MODEL,
+      messages: [{
+        role: 'user',
+        content,
+      }],
+      max_tokens: 4096 * batch.length,
+      temperature: 0.0,
+    }),
+    signal: AbortSignal.timeout(600000),
+  });
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+
+  if (!response.ok) {
+    const errorText = await response.text();
+    throw new Error(`Nanonets API error: ${response.status} - ${errorText}`);
+  }
+
+  const data = await response.json();
+  let responseContent = (data.choices?.[0]?.message?.content || '').trim();
+
+  if (batch.length === 1 && !responseContent.includes('--- PAGE')) {
+    responseContent = `--- PAGE ${batch[0].pageNum} ---\n${responseContent}`;
+  }
+
+  console.log(`        Pages [${pageNums}]: ${responseContent.length} chars (${elapsed}s)`);
+  return responseContent;
+}
+
+/**
+ * Convert all pages of a document to markdown using smart batching
+ */
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
+  const batches = batchImages(images);
+  console.log(`      [${docName}] Processing ${images.length} page(s) in ${batches.length} batch(es)...`);
+
+  const markdownParts: string[] = [];
+
+  for (let i = 0; i < batches.length; i++) {
+    const batch = batches[i];
+    const batchTokens = batch.reduce((sum, img) => sum + estimateVisualTokens(img.width, img.height), 0);
+    console.log(`        Batch ${i + 1}: ${batch.length} page(s), ~${batchTokens} tokens`);
+    const markdown = await convertBatchToMarkdown(batch);
+    markdownParts.push(markdown);
+  }
+
+  const fullMarkdown = markdownParts.join('\n\n');
+  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
+  return fullMarkdown;
+}
+
+/**
+ * Stop Nanonets container
+ */
+function stopNanonets(): void {
+  console.log('    [Docker] Stopping Nanonets container...');
+  try {
+    execSync('docker stop nanonets-test 2>/dev/null || true', { stdio: 'pipe' });
+    execSync('sleep 5', { stdio: 'pipe' });
+    console.log('    [Docker] Nanonets stopped');
+  } catch {
+    console.log('    [Docker] Nanonets was not running');
+  }
+}
+
+/**
+ * Ensure GPT-OSS 20B model is available
+ */
+async function ensureExtractionModel(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_URL}/api/tags`);
+    if (response.ok) {
+      const data = await response.json();
+      const models = data.models || [];
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`    [Ollama] Pulling ${EXTRACTION_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: EXTRACTION_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+/**
+ * Parse amount from string (handles European format)
+ */
+function parseAmount(s: string | number | undefined): number {
+  if (s === undefined || s === null) return 0;
+  if (typeof s === 'number') return s;
+  const match = s.match(/([\d.,]+)/);
+  if (!match) return 0;
+  const numStr = match[1];
+  const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
+    ? numStr.replace(/\./g, '').replace(',', '.')
+    : numStr.replace(/,/g, '');
+  return parseFloat(normalized) || 0;
+}
+
+/**
+ * Extract invoice number - MINIMAL normalization for debugging
+ */
+function extractInvoiceNumber(s: string | undefined): string {
+  if (!s) return '';
+  // Only remove markdown formatting, preserve everything else
+  return s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+}
+
+/**
+ * Extract date (YYYY-MM-DD) from response
+ */
+function extractDate(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
+  if (isoMatch) return isoMatch[1];
+  const dmyMatch = clean.match(/(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})/);
+  if (dmyMatch) {
+    return `${dmyMatch[3]}-${dmyMatch[2].padStart(2, '0')}-${dmyMatch[1].padStart(2, '0')}`;
+  }
+  return clean.replace(/[^\d-]/g, '').trim();
+}
+
+/**
+ * Extract currency
+ */
+function extractCurrency(s: string | undefined): string {
+  if (!s) return 'EUR';
+  const upper = s.toUpperCase();
+  if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
+  if (upper.includes('USD') || upper.includes('$')) return 'USD';
+  if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
+  return 'EUR';
+}
+
+/**
+ * Extract JSON from response
+ */
+function extractJsonFromResponse(response: string): Record<string, unknown> | null {
+  let cleanResponse = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+  const codeBlockMatch = cleanResponse.match(/```(?:json)?\s*([\s\S]*?)```/);
+  const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : cleanResponse;
+
+  try {
+    return JSON.parse(jsonStr);
+  } catch {
+    const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
+    if (jsonMatch) {
+      try {
+        return JSON.parse(jsonMatch[0]);
+      } catch {
+        return null;
+      }
+    }
+    return null;
+  }
+}
+
+/**
+ * Parse JSON response into IInvoice
+ */
+function parseJsonToInvoice(response: string): IInvoice | null {
+  const parsed = extractJsonFromResponse(response);
+  if (!parsed) return null;
+
+  return {
+    invoice_number: extractInvoiceNumber(String(parsed.invoice_number || '')),
+    invoice_date: extractDate(String(parsed.invoice_date || '')),
+    vendor_name: String(parsed.vendor_name || '').replace(/\*\*/g, '').replace(/`/g, '').trim(),
+    currency: extractCurrency(String(parsed.currency || '')),
+    net_amount: parseAmount(parsed.net_amount as string | number),
+    vat_amount: parseAmount(parsed.vat_amount as string | number),
+    total_amount: parseAmount(parsed.total_amount as string | number),
+  };
+}
+
+/**
+ * Extract invoice from markdown using GPT-OSS 20B (streaming)
+ */
+async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Promise<IInvoice | null> {
+  const startTime = Date.now();
+  const fullPrompt = JSON_EXTRACTION_PROMPT + markdown;
+
+  // Log exact prompt
+  console.log(`\n      [${queryId}] ===== PROMPT =====`);
+  console.log(fullPrompt);
+  console.log(`      [${queryId}] ===== END PROMPT (${fullPrompt.length} chars) =====\n`);
+
+  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: EXTRACTION_MODEL,
+      messages: [
+        { role: 'user', content: 'Hi there, how are you?' },
+        { role: 'assistant', content: 'Good, how can I help you today?' },
+        { role: 'user', content: fullPrompt },
+      ],
+      stream: true,
+    }),
+    signal: AbortSignal.timeout(600000),
+  });
+
+  if (!response.ok) {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] ERROR: ${response.status} (${elapsed}s)`);
+    throw new Error(`Ollama API error: ${response.status}`);
+  }
+
+  // Stream the response
+  let content = '';
+  let thinkingContent = '';
+  let thinkingStarted = false;
+  let outputStarted = false;
+  const reader = response.body!.getReader();
+  const decoder = new TextDecoder();
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+
+      const chunk = decoder.decode(value, { stream: true });
+
+      for (const line of chunk.split('\n').filter(l => l.trim())) {
+        try {
+          const json = JSON.parse(line);
+
+          const thinking = json.message?.thinking || '';
+          if (thinking) {
+            if (!thinkingStarted) {
+              process.stdout.write(`      [${queryId}] THINKING: `);
+              thinkingStarted = true;
+            }
+            process.stdout.write(thinking);
+            thinkingContent += thinking;
+          }
+
+          const token = json.message?.content || '';
+          if (token) {
+            if (!outputStarted) {
+              if (thinkingStarted) process.stdout.write('\n');
+              process.stdout.write(`      [${queryId}] OUTPUT: `);
+              outputStarted = true;
+            }
+            process.stdout.write(token);
+            content += token;
+          }
+        } catch {
+          // Ignore parse errors for partial chunks
+        }
+      }
+    }
+  } finally {
+    if (thinkingStarted || outputStarted) process.stdout.write('\n');
+  }
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  console.log(`      [${queryId}] Done: ${thinkingContent.length} thinking chars, ${content.length} output chars (${elapsed}s)`);
+
+  // Log raw response for debugging
+  console.log(`      [${queryId}] RAW RESPONSE: ${content}`);
+
+  return parseJsonToInvoice(content);
+}
+
+/**
+ * Extract invoice (single pass)
+ */
+async function extractInvoice(markdown: string, docName: string): Promise<IInvoice> {
+  console.log(`    [${docName}] Extracting...`);
+  const invoice = await extractInvoiceFromMarkdown(markdown, docName);
+  if (!invoice) {
+    return {
+      invoice_number: '',
+      invoice_date: '',
+      vendor_name: '',
+      currency: 'EUR',
+      net_amount: 0,
+      vat_amount: 0,
+      total_amount: 0,
+    };
+  }
+  console.log(`    [${docName}] Extracted: ${JSON.stringify(invoice, null, 2)}`);
+  return invoice;
+}
+
+/**
+ * Normalize date to YYYY-MM-DD
+ */
+function normalizeDate(dateStr: string | null): string {
+  if (!dateStr) return '';
+  if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
+
+  const monthMap: Record<string, string> = {
+    JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
+    JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
+  };
+
+  let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
+  if (match) {
+    return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
+  }
+
+  match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
+  if (match) {
+    return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
+  }
+
+  return dateStr;
+}
+
+/**
+ * Compare extracted invoice against expected - detailed output
+ */
+function compareInvoice(
+  extracted: IInvoice,
+  expected: IInvoice
+): { match: boolean; errors: string[] } {
+  const errors: string[] = [];
+
+  // Invoice number comparison - exact match after whitespace normalization
+  const extNum = extracted.invoice_number?.trim() || '';
+  const expNum = expected.invoice_number?.trim() || '';
+  if (extNum.toLowerCase() !== expNum.toLowerCase()) {
+    errors.push(`invoice_number: expected "${expected.invoice_number}", got "${extracted.invoice_number}"`);
+  }
+
+  if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
+    errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
+  }
+
+  if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
+    errors.push(`total_amount: expected ${expected.total_amount}, got ${extracted.total_amount}`);
+  }
+
+  if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
+    errors.push(`currency: expected "${expected.currency}", got "${extracted.currency}"`);
+  }
+
+  return { match: errors.length === 0, errors };
+}
+
+/**
+ * Find test cases for failed invoices only
+ */
+function findTestCases(): ITestCase[] {
+  const testDir = path.join(process.cwd(), '.nogit/invoices');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: ITestCase[] = [];
+
+  for (const invoiceName of FAILED_INVOICES) {
+    const pdfFile = `${invoiceName}.pdf`;
+    const jsonFile = `${invoiceName}.json`;
+
+    if (files.includes(pdfFile) && files.includes(jsonFile)) {
+      testCases.push({
+        name: invoiceName,
+        pdfPath: path.join(testDir, pdfFile),
+        jsonPath: path.join(testDir, jsonFile),
+      });
+    } else {
+      console.warn(`Warning: Missing files for ${invoiceName}`);
+    }
+  }
+
+  return testCases;
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\n========================================`);
+console.log(`  FAILED INVOICES DEBUG TEST`);
+console.log(`========================================`);
+console.log(`  Testing ${testCases.length} failed invoices:`);
+for (const tc of testCases) {
+  console.log(`    - ${tc.name}`);
+}
+console.log(`========================================\n`);
+
+// Ensure temp directory exists
+if (!fs.existsSync(TEMP_MD_DIR)) {
+  fs.mkdirSync(TEMP_MD_DIR, { recursive: true });
+}
+
+// -------- STAGE 1: OCR with Nanonets --------
+
+tap.test('Stage 1: Setup Nanonets', async () => {
+  console.log('\n========== STAGE 1: Nanonets OCR ==========\n');
+  const ok = await ensureNanonetsOcr();
+  expect(ok).toBeTrue();
+});
+
+tap.test('Stage 1: Convert failed invoices to markdown', async () => {
+  console.log('\n  Converting failed invoice PDFs to markdown with Nanonets-OCR-s...\n');
+
+  for (const tc of testCases) {
+    console.log(`\n    === ${tc.name} ===`);
+
+    const images = convertPdfToImages(tc.pdfPath);
+    console.log(`      Pages: ${images.length}`);
+
+    const markdown = await convertDocumentToMarkdown(images, tc.name);
+
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    fs.writeFileSync(mdPath, markdown);
+    tc.markdownPath = mdPath;
+    console.log(`      Saved: ${mdPath}`);
+
+    // Also save to .nogit for inspection
+    const debugMdPath = path.join(process.cwd(), '.nogit/invoices', `${tc.name}.debug.md`);
+    fs.writeFileSync(debugMdPath, markdown);
+    console.log(`      Debug copy: ${debugMdPath}`);
+  }
+
+  console.log('\n  Stage 1 complete: All failed invoices converted to markdown\n');
+});
+
+tap.test('Stage 1: Stop Nanonets', async () => {
+  stopNanonets();
+  await new Promise(resolve => setTimeout(resolve, 3000));
+  expect(isContainerRunning('nanonets-test')).toBeFalse();
+});
+
+// -------- STAGE 2: Extraction with GPT-OSS 20B --------
+
+tap.test('Stage 2: Setup Ollama + GPT-OSS 20B', async () => {
+  console.log('\n========== STAGE 2: GPT-OSS 20B Extraction ==========\n');
+
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
+});
+
+let passedCount = 0;
+let failedCount = 0;
+
+for (const tc of testCases) {
+  tap.test(`Stage 2: Extract ${tc.name}`, async () => {
+    const expected: IInvoice = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    console.log(`\n    ========================================`);
+    console.log(`    === ${tc.name} ===`);
+    console.log(`    ========================================`);
+    console.log(`    EXPECTED:`);
+    console.log(`      invoice_number: "${expected.invoice_number}"`);
+    console.log(`      invoice_date:   "${expected.invoice_date}"`);
+    console.log(`      vendor_name:    "${expected.vendor_name}"`);
+    console.log(`      total_amount:   ${expected.total_amount} ${expected.currency}`);
+
+    const startTime = Date.now();
+
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      throw new Error(`Markdown not found: ${mdPath}. Run Stage 1 first.`);
+    }
+    const markdown = fs.readFileSync(mdPath, 'utf-8');
+    console.log(`    Markdown: ${markdown.length} chars`);
+
+    const extracted = await extractInvoice(markdown, tc.name);
+
+    const elapsedMs = Date.now() - startTime;
+
+    console.log(`\n    EXTRACTED:`);
+    console.log(`      invoice_number: "${extracted.invoice_number}"`);
+    console.log(`      invoice_date:   "${extracted.invoice_date}"`);
+    console.log(`      vendor_name:    "${extracted.vendor_name}"`);
+    console.log(`      total_amount:   ${extracted.total_amount} ${extracted.currency}`);
+
+    const result = compareInvoice(extracted, expected);
+
+    if (result.match) {
+      passedCount++;
+      console.log(`\n    Result: ✓ MATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+    } else {
+      failedCount++;
+      console.log(`\n    Result: ✗ MISMATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+      console.log(`    ERRORS:`);
+      result.errors.forEach(e => console.log(`      - ${e}`));
+    }
+
+    // Don't fail the test - we're debugging
+    // expect(result.match).toBeTrue();
+  });
+}
+
+tap.test('Summary', async () => {
+  const totalInvoices = testCases.length;
+  const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
+
+  console.log(`\n========================================`);
+  console.log(`   Failed Invoices Debug Summary`);
+  console.log(`========================================`);
+  console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
+  console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
+  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
+  console.log(`========================================`);
+  console.log(`  Markdown files saved to: ${TEMP_MD_DIR}`);
+  console.log(`  Debug copies in: .nogit/invoices/*.debug.md`);
+  console.log(`========================================\n`);
+
+  // Don't cleanup temp files for debugging
+  console.log(`  Keeping temp files for debugging.\n`);
+});
+
+export default tap.start();

test/test.invoices.minicpm.ts

@@ -1,10 +1,10 @@
 /**
- * Invoice extraction test using MiniCPM-V (visual extraction)
+ * Invoice extraction test using MiniCPM-V via smartagent DualAgentOrchestrator
  *
- * Consensus approach:
- * 1. Pass 1: Fast JSON extraction
- * 2. Pass 2: Confirm with thinking enabled
- * 3. If mismatch: repeat until consensus or max attempts
+ * Uses vision-capable orchestrator with JsonValidatorTool for self-validation:
+ * 1. Pass images to the orchestrator
+ * 2. Driver extracts invoice data and validates JSON before completing
+ * 3. If validation fails, driver retries within the same task
  */
 import { tap, expect } from '@git.zone/tstest/tapbundle';
 import * as fs from 'fs';
@@ -12,6 +12,8 @@ import * as path from 'path';
 import { execSync } from 'child_process';
 import * as os from 'os';
 import { ensureMiniCpm } from './helpers/docker.js';
+import { SmartAi } from '@push.rocks/smartai';
+import { DualAgentOrchestrator, JsonValidatorTool } from '@push.rocks/smartagent';
 
 const OLLAMA_URL = 'http://localhost:11434';
 const MODEL = 'openbmb/minicpm-v4.5:q8_0';
@@ -26,6 +28,10 @@ interface IInvoice {
   total_amount: number;
 }
 
+// SmartAi instance and orchestrator (initialized in setup)
+let smartAi: SmartAi;
+let orchestrator: DualAgentOrchestrator;
+
 /**
  * Convert PDF to PNG images using ImageMagick
  */
@@ -54,7 +60,9 @@ function convertPdfToImages(pdfPath: string): string[] {
   }
 }
 
-const JSON_PROMPT = `Extract invoice data from this image. Return ONLY a JSON object with these exact fields:
+const EXTRACTION_PROMPT = `Extract invoice data from the provided image(s).
+
+IMPORTANT: You must output a valid JSON object with these exact fields:
 {
   "invoice_number": "the invoice number (not VAT ID, not customer ID)",
   "invoice_date": "YYYY-MM-DD format",
@@ -64,67 +72,16 @@ const JSON_PROMPT = `Extract invoice data from this image. Return ONLY a JSON ob
   "vat_amount": 0.00,
   "total_amount": 0.00
 }
-Return only the JSON, no explanation.`;
 
-/**
- * Query MiniCPM-V for JSON output (fast, no thinking)
- */
-async function queryJsonFast(images: string[]): Promise<string> {
-  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      model: MODEL,
-      messages: [{
-        role: 'user',
-        content: JSON_PROMPT,
-        images: images,
-      }],
-      stream: false,
-      options: {
-        num_predict: 1000,
-        temperature: 0.1,
-      },
-    }),
-  });
+Before completing, use the json.validate tool to verify your output is valid JSON with all required fields.
 
-  if (!response.ok) {
-    throw new Error(`Ollama API error: ${response.status}`);
-  }
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON_HERE", "requiredFields": ["invoice_number", "invoice_date", "vendor_name", "currency", "net_amount", "vat_amount", "total_amount"]}</params>
+</tool_call>
 
-  const data = await response.json();
-  return (data.message?.content || '').trim();
-}
-
-/**
- * Query MiniCPM-V for JSON output with thinking enabled (slower, more accurate)
- */
-async function queryJsonWithThinking(images: string[]): Promise<string> {
-  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      model: MODEL,
-      messages: [{
-        role: 'user',
-        content: `Think carefully about this invoice image, then ${JSON_PROMPT}`,
-        images: images,
-      }],
-      stream: false,
-      options: {
-        num_predict: 2000,
-        temperature: 0.1,
-      },
-    }),
-  });
-
-  if (!response.ok) {
-    throw new Error(`Ollama API error: ${response.status}`);
-  }
-
-  const data = await response.json();
-  return (data.message?.content || '').trim();
-}
+Only complete the task after validation passes. Output the final JSON in <task_complete> tags.`;
 
 /**
  * Parse amount from string (handles European format)
@@ -190,9 +147,31 @@ function extractCurrency(s: string | undefined): string {
 }
 
 /**
- * Extract JSON from response (handles markdown code blocks)
+ * Extract JSON from response (handles markdown code blocks and task_complete tags)
  */
 function extractJsonFromResponse(response: string): Record<string, unknown> | null {
+  // Try to find JSON in task_complete tags
+  const completeMatch = response.match(/<task_complete>([\s\S]*?)<\/task_complete>/);
+  if (completeMatch) {
+    const content = completeMatch[1].trim();
+    // Try to find JSON in the content
+    const codeBlockMatch = content.match(/```(?:json)?\s*([\s\S]*?)```/);
+    const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : content;
+    try {
+      return JSON.parse(jsonStr);
+    } catch {
+      // Try to find JSON object pattern
+      const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
+      if (jsonMatch) {
+        try {
+          return JSON.parse(jsonMatch[0]);
+        } catch {
+          return null;
+        }
+      }
+    }
+  }
+
   // Try to find JSON in markdown code block
   const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
   const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
@@ -232,76 +211,27 @@ function parseJsonToInvoice(response: string): IInvoice | null {
 }
 
 /**
- * Compare two invoices for consensus (key fields must match)
- */
-function invoicesMatch(a: IInvoice, b: IInvoice): boolean {
-  const numMatch = a.invoice_number.toLowerCase() === b.invoice_number.toLowerCase();
-  const dateMatch = a.invoice_date === b.invoice_date;
-  const totalMatch = Math.abs(a.total_amount - b.total_amount) < 0.02;
-  return numMatch && dateMatch && totalMatch;
-}
-
-/**
- * Extract invoice data using consensus approach:
- * 1. Pass 1: Fast JSON extraction
- * 2. Pass 2: Confirm with thinking enabled
- * 3. If mismatch: repeat until consensus or max 5 attempts
+ * Extract invoice data using smartagent orchestrator with vision
  */
 async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
-  console.log(`    [Vision] Processing ${images.length} page(s) with ${MODEL} (consensus)`);
+  console.log(`    [Vision] Processing ${images.length} page(s) with smartagent DualAgentOrchestrator`);
 
-  const MAX_ATTEMPTS = 5;
-  let attempt = 0;
+  const startTime = Date.now();
 
-  while (attempt < MAX_ATTEMPTS) {
-    attempt++;
-    console.log(`    [Attempt ${attempt}/${MAX_ATTEMPTS}]`);
+  const result = await orchestrator.run(EXTRACTION_PROMPT, { images });
 
-    // PASS 1: Fast JSON extraction
-    console.log(`    [Pass 1] Fast extraction...`);
-    const fastResponse = await queryJsonFast(images);
-    const fastInvoice = parseJsonToInvoice(fastResponse);
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  console.log(`    [Vision] Completed in ${elapsed}s (${result.iterations} iterations, status: ${result.status})`);
 
-    if (!fastInvoice) {
-      console.log(`    [Pass 1] JSON parsing failed, retrying...`);
-      continue;
-    }
-    console.log(`    [Pass 1] Result: ${fastInvoice.invoice_number} | ${fastInvoice.invoice_date} | ${fastInvoice.total_amount} ${fastInvoice.currency}`);
+  const invoice = parseJsonToInvoice(result.result);
 
-    // PASS 2: Confirm with thinking
-    console.log(`    [Pass 2] Thinking confirmation...`);
-    const thinkResponse = await queryJsonWithThinking(images);
-    const thinkInvoice = parseJsonToInvoice(thinkResponse);
-
-    if (!thinkInvoice) {
-      console.log(`    [Pass 2] JSON parsing failed, retrying...`);
-      continue;
-    }
-    console.log(`    [Pass 2] Result: ${thinkInvoice.invoice_number} | ${thinkInvoice.invoice_date} | ${thinkInvoice.total_amount} ${thinkInvoice.currency}`);
-
-    // Check consensus
-    if (invoicesMatch(fastInvoice, thinkInvoice)) {
-      console.log(`    [Consensus] MATCH - using result`);
-      return thinkInvoice; // Prefer thinking result
-    }
-
-    console.log(`    [Consensus] MISMATCH - repeating...`);
-    console.log(`      Fast:  ${fastInvoice.invoice_number} | ${fastInvoice.invoice_date} | ${fastInvoice.total_amount}`);
-    console.log(`      Think: ${thinkInvoice.invoice_number} | ${thinkInvoice.invoice_date} | ${thinkInvoice.total_amount}`);
+  if (invoice) {
+    console.log(`    [Result] ${invoice.invoice_number} | ${invoice.invoice_date} | ${invoice.total_amount} ${invoice.currency}`);
+    return invoice;
   }
 
-  // Max attempts reached - do one final thinking pass and use that
-  console.log(`    [Final] Max attempts reached, using final thinking pass`);
-  const finalResponse = await queryJsonWithThinking(images);
-  const finalInvoice = parseJsonToInvoice(finalResponse);
-
-  if (finalInvoice) {
-    console.log(`    [Final] Result: ${finalInvoice.invoice_number} | ${finalInvoice.invoice_date} | ${finalInvoice.total_amount} ${finalInvoice.currency}`);
-    return finalInvoice;
-  }
-
-  // Return empty invoice if all else fails
-  console.log(`    [Final] All parsing failed, returning empty`);
+  // Return empty invoice if parsing failed
+  console.log(`    [Result] Parsing failed, returning empty invoice`);
   return {
     invoice_number: '',
     invoice_date: '',
@@ -410,6 +340,79 @@ tap.test('setup: ensure Docker containers are running', async () => {
   console.log('\n[Setup] All containers ready!\n');
 });
 
+tap.test('setup: initialize smartagent orchestrator', async () => {
+  console.log('[Setup] Initializing SmartAi and DualAgentOrchestrator...');
+
+  smartAi = new SmartAi({
+    ollama: {
+      baseUrl: OLLAMA_URL,
+      model: MODEL,
+      defaultOptions: {
+        num_ctx: 32768,
+        temperature: 0.1,
+      },
+      defaultTimeout: 300000, // 5 minutes for vision tasks
+    },
+  });
+
+  await smartAi.start();
+
+  orchestrator = new DualAgentOrchestrator({
+    smartAiInstance: smartAi,
+    defaultProvider: 'ollama',
+    guardianPolicyPrompt: `You are a Guardian agent overseeing invoice extraction tasks.
+
+APPROVE all tool calls that:
+- Use the json.validate action to verify JSON output
+- Are reasonable attempts to complete the extraction task
+
+REJECT tool calls that:
+- Attempt to access external resources
+- Try to execute arbitrary code
+- Are clearly unrelated to invoice extraction`,
+    driverSystemMessage: `You are an AI assistant that extracts invoice data from images.
+
+Your task is to analyze invoice images and extract structured data.
+You have access to a json.validate tool to verify your JSON output.
+
+IMPORTANT: Always validate your JSON before completing the task.
+
+## Tool Usage Format
+When you need to validate JSON, output:
+
+<tool_call>
+  <tool>json</tool>
+  <action>validate</action>
+  <params>{"jsonString": "YOUR_JSON", "requiredFields": ["invoice_number", "invoice_date", "vendor_name", "currency", "net_amount", "vat_amount", "total_amount"]}</params>
+</tool_call>
+
+## Completion Format
+After validation passes, complete the task:
+
+<task_complete>
+{"invoice_number": "...", "invoice_date": "YYYY-MM-DD", ...}
+</task_complete>`,
+    maxIterations: 5,
+    maxConsecutiveRejections: 3,
+    onToken: (token, source) => {
+      if (source === 'driver') {
+        process.stdout.write(token);
+      }
+    },
+    onProgress: (event) => {
+      if (event.logLevel === 'error') {
+        console.error(event.logMessage);
+      }
+    },
+  });
+
+  // Register the JsonValidatorTool
+  orchestrator.registerTool(new JsonValidatorTool());
+
+  await orchestrator.start();
+  console.log('[Setup] Orchestrator initialized!\n');
+});
+
 tap.test('should have MiniCPM-V model loaded', async () => {
   const response = await fetch(`${OLLAMA_URL}/api/tags`);
   const data = await response.json();
@@ -418,7 +421,7 @@ tap.test('should have MiniCPM-V model loaded', async () => {
 });
 
 const testCases = findTestCases();
-console.log(`\nFound ${testCases.length} invoice test cases (MiniCPM-V)\n`);
+console.log(`\nFound ${testCases.length} invoice test cases (smartagent + MiniCPM-V)\n`);
 
 let passedCount = 0;
 let failedCount = 0;
@@ -455,6 +458,13 @@ for (const testCase of testCases) {
   });
 }
 
+tap.test('cleanup: stop orchestrator', async () => {
+  if (orchestrator) {
+    await orchestrator.stop();
+  }
+  console.log('[Cleanup] Orchestrator stopped');
+});
+
 tap.test('summary', async () => {
   const totalInvoices = testCases.length;
   const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
@@ -462,9 +472,10 @@ tap.test('summary', async () => {
   const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
 
   console.log(`\n========================================`);
-  console.log(`   Invoice Extraction Summary (${MODEL})`);
+  console.log(`   Invoice Extraction Summary`);
+  console.log(`   (smartagent + ${MODEL})`);
   console.log(`========================================`);
-  console.log(`  Method:    Consensus (fast + thinking)`);
+  console.log(`  Method:    DualAgentOrchestrator with vision`);
   console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
   console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
   console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);

test/test.invoices.nanonets.ts

@@ -0,0 +1,752 @@
+/**
+ * Invoice extraction using Nanonets-OCR2-3B + GPT-OSS 20B (sequential two-stage pipeline)
+ *
+ * Stage 1: Nanonets-OCR2-3B converts ALL document pages to markdown (stop after completion)
+ * Stage 2: GPT-OSS 20B extracts structured JSON from saved markdown (after Nanonets stops)
+ *
+ * This approach avoids GPU contention by running services sequentially.
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { execSync } from 'child_process';
+import * as os from 'os';
+import { ensureNanonetsOcr, ensureMiniCpm, isContainerRunning } from './helpers/docker.js';
+import { SmartAi } from '@push.rocks/smartai';
+import { DualAgentOrchestrator, JsonValidatorTool } from '@push.rocks/smartagent';
+
+const NANONETS_URL = 'http://localhost:8000/v1';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR2-3B';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const EXTRACTION_MODEL = 'gpt-oss:20b';
+
+// Persistent cache directory for storing markdown between runs
+const MD_CACHE_DIR = path.join(process.cwd(), '.nogit/invoices-md');
+
+// SmartAi instance for Ollama with optimized settings
+const smartAi = new SmartAi({
+  ollama: {
+    baseUrl: OLLAMA_URL,
+    model: EXTRACTION_MODEL,
+    defaultOptions: {
+      num_ctx: 65536,    // 64K context for long invoices + reasoning chains
+      temperature: 0,    // Deterministic for JSON extraction
+      repeat_penalty: 1.3, // Penalty to prevent repetition loops
+      think: true,       // Enable thinking mode for GPT-OSS reasoning
+    },
+    defaultTimeout: 600000,  // 10 minute timeout for large documents
+  },
+});
+
+// DualAgentOrchestrator for structured task execution
+let orchestrator: DualAgentOrchestrator;
+
+interface IInvoice {
+  invoice_number: string;
+  invoice_date: string;
+  vendor_name: string;
+  currency: string;
+  net_amount: number;
+  vat_amount: number;
+  total_amount: number;
+}
+
+interface IImageData {
+  base64: string;
+  width: number;
+  height: number;
+  pageNum: number;
+}
+
+interface ITestCase {
+  name: string;
+  pdfPath: string;
+  jsonPath: string;
+  markdownPath?: string;
+}
+
+// Nanonets-specific prompt for document OCR to markdown
+const NANONETS_OCR_PROMPT = `Extract the text from the above document as if you were reading it naturally.
+Return the tables in html format.
+Return the equations in LaTeX representation.
+If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
+Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
+Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.`;
+
+// JSON extraction prompt for GPT-OSS 20B (sent AFTER the invoice text is provided)
+const JSON_EXTRACTION_PROMPT = `Extract key fields from the invoice. Return ONLY valid JSON.
+
+WHERE TO FIND DATA:
+- invoice_number, invoice_date, vendor_name: Look in the HEADER section at the TOP of PAGE 1 (near "Invoice no.", "Invoice date:", "Rechnungsnummer"). Use common sense. Btw. an invoice number might start on INV* .
+- net_amount, vat_amount, total_amount: Look in the SUMMARY section at the BOTTOM (look for "Total", "Amount due", "Gesamtbetrag")
+
+RULES:
+1. Use common sense.
+2. invoice_date: Convert to YYYY-MM-DD format (e.g., "14/04/2022" → "2022-04-14")
+3. vendor_name: The company issuing the invoice
+4. currency: EUR, USD, or GBP
+5. net_amount: Total before tax
+6. vat_amount: Tax amount
+7. total_amount: Final total with tax
+
+JSON only:
+{"invoice_number":"X","invoice_date":"YYYY-MM-DD","vendor_name":"X","currency":"EUR","net_amount":0,"vat_amount":0,"total_amount":0}
+
+Double check for valid JSON syntax.
+
+`;
+
+// Constants for smart batching
+const PATCH_SIZE = 14;  // Qwen2.5-VL uses 14x14 patches
+
+/**
+ * Estimate visual tokens for an image based on dimensions
+ */
+function estimateVisualTokens(width: number, height: number): number {
+  return Math.ceil((width * height) / (PATCH_SIZE * PATCH_SIZE));
+}
+
+/**
+ * Process images one page at a time for reliability
+ */
+function batchImages(images: IImageData[]): IImageData[][] {
+  // One page per batch for reliable processing
+  return images.map(img => [img]);
+}
+
+/**
+ * Convert PDF to JPEG images using ImageMagick with dimension tracking
+ */
+function convertPdfToImages(pdfPath: string): IImageData[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
+  const outputPattern = path.join(tempDir, 'page-%d.jpg');
+
+  try {
+    execSync(
+      `convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
+      { stdio: 'pipe' }
+    );
+
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
+    const images: IImageData[] = [];
+
+    for (let i = 0; i < files.length; i++) {
+      const file = files[i];
+      const imagePath = path.join(tempDir, file);
+      const imageData = fs.readFileSync(imagePath);
+
+      // Get image dimensions using identify command
+      const dimensions = execSync(`identify -format "%w %h" "${imagePath}"`, { encoding: 'utf-8' }).trim();
+      const [width, height] = dimensions.split(' ').map(Number);
+
+      images.push({
+        base64: imageData.toString('base64'),
+        width,
+        height,
+        pageNum: i + 1,
+      });
+    }
+
+    return images;
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Convert a batch of pages to markdown using Nanonets-OCR-s
+ */
+async function convertBatchToMarkdown(batch: IImageData[]): Promise<string> {
+  const startTime = Date.now();
+  const pageNums = batch.map(img => img.pageNum).join(', ');
+
+  // Build content array with all images first, then the prompt
+  const content: Array<{ type: string; image_url?: { url: string }; text?: string }> = [];
+
+  for (const img of batch) {
+    content.push({
+      type: 'image_url',
+      image_url: { url: `data:image/jpeg;base64,${img.base64}` },
+    });
+  }
+
+  // Add prompt with page separator instruction if multiple pages
+  const promptText = batch.length > 1
+    ? `${NANONETS_OCR_PROMPT}\n\nPlease clearly separate each page's content with "--- PAGE N ---" markers, where N is the page number starting from ${batch[0].pageNum}.`
+    : NANONETS_OCR_PROMPT;
+
+  content.push({ type: 'text', text: promptText });
+
+  const response = await fetch(`${NANONETS_URL}/chat/completions`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': 'Bearer dummy',
+    },
+    body: JSON.stringify({
+      model: NANONETS_MODEL,
+      messages: [{
+        role: 'user',
+        content,
+      }],
+      max_tokens: 4096 * batch.length,  // Scale output tokens with batch size
+      temperature: 0.0,
+    }),
+    signal: AbortSignal.timeout(600000),  // 10 minute timeout for OCR
+  });
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+
+  if (!response.ok) {
+    const errorText = await response.text();
+    throw new Error(`Nanonets API error: ${response.status} - ${errorText}`);
+  }
+
+  const data = await response.json();
+  let responseContent = (data.choices?.[0]?.message?.content || '').trim();
+
+  // For single-page batches, add page marker if not present
+  if (batch.length === 1 && !responseContent.includes('--- PAGE')) {
+    responseContent = `--- PAGE ${batch[0].pageNum} ---\n${responseContent}`;
+  }
+
+  console.log(`        Pages [${pageNums}]: ${responseContent.length} chars (${elapsed}s)`);
+  return responseContent;
+}
+
+/**
+ * Convert all pages of a document to markdown using smart batching
+ */
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
+  const batches = batchImages(images);
+  console.log(`      [${docName}] Processing ${images.length} page(s) in ${batches.length} batch(es)...`);
+
+  const markdownParts: string[] = [];
+
+  for (let i = 0; i < batches.length; i++) {
+    const batch = batches[i];
+    const batchTokens = batch.reduce((sum, img) => sum + estimateVisualTokens(img.width, img.height), 0);
+    console.log(`        Batch ${i + 1}: ${batch.length} page(s), ~${batchTokens} tokens`);
+    const markdown = await convertBatchToMarkdown(batch);
+    markdownParts.push(markdown);
+  }
+
+  const fullMarkdown = markdownParts.join('\n\n');
+  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
+  return fullMarkdown;
+}
+
+/**
+ * Stop Nanonets container
+ */
+function stopNanonets(): void {
+  console.log('    [Docker] Stopping Nanonets container...');
+  try {
+    execSync('docker stop nanonets-test 2>/dev/null || true', { stdio: 'pipe' });
+    execSync('sleep 5', { stdio: 'pipe' });
+    console.log('    [Docker] Nanonets stopped');
+  } catch {
+    console.log('    [Docker] Nanonets was not running');
+  }
+}
+
+/**
+ * Ensure GPT-OSS 20B model is available
+ */
+async function ensureExtractionModel(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_URL}/api/tags`);
+    if (response.ok) {
+      const data = await response.json();
+      const models = data.models || [];
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`    [Ollama] Pulling ${EXTRACTION_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: EXTRACTION_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+/**
+ * Parse amount from string (handles European format)
+ */
+function parseAmount(s: string | number | undefined): number {
+  if (s === undefined || s === null) return 0;
+  if (typeof s === 'number') return s;
+  const match = s.match(/([\d.,]+)/);
+  if (!match) return 0;
+  const numStr = match[1];
+  const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
+    ? numStr.replace(/\./g, '').replace(',', '.')
+    : numStr.replace(/,/g, '');
+  return parseFloat(normalized) || 0;
+}
+
+/**
+ * Extract invoice number from potentially verbose response
+ */
+function extractInvoiceNumber(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const patterns = [
+    /\b([A-Z]{2,3}\d{10,})\b/i,
+    /\b([A-Z]\d{8,})\b/i,
+    /\b(INV[-\s]?\d{4}[-\s]?\d+)\b/i,
+    /\b(\d{7,})\b/,
+  ];
+  for (const pattern of patterns) {
+    const match = clean.match(pattern);
+    if (match) return match[1];
+  }
+  return clean.replace(/[^A-Z0-9-]/gi, '').trim() || clean;
+}
+
+/**
+ * Extract date (YYYY-MM-DD) from response
+ */
+function extractDate(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
+  if (isoMatch) return isoMatch[1];
+  const dmyMatch = clean.match(/(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})/);
+  if (dmyMatch) {
+    return `${dmyMatch[3]}-${dmyMatch[2].padStart(2, '0')}-${dmyMatch[1].padStart(2, '0')}`;
+  }
+  return clean.replace(/[^\d-]/g, '').trim();
+}
+
+/**
+ * Extract currency
+ */
+function extractCurrency(s: string | undefined): string {
+  if (!s) return 'EUR';
+  const upper = s.toUpperCase();
+  if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
+  if (upper.includes('USD') || upper.includes('$')) return 'USD';
+  if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
+  return 'EUR';
+}
+
+/**
+ * Try to extract valid JSON from a response string
+ */
+function tryExtractJson(response: string): Record<string, unknown> | null {
+  // Remove thinking tags
+  let clean = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+
+  // Try code block
+  const codeBlockMatch = clean.match(/```(?:json)?\s*([\s\S]*?)```/);
+  const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : clean;
+
+  try {
+    return JSON.parse(jsonStr);
+  } catch {
+    // Try to find JSON object
+    const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
+    if (jsonMatch) {
+      try {
+        return JSON.parse(jsonMatch[0]);
+      } catch {
+        return null;
+      }
+    }
+    return null;
+  }
+}
+
+/**
+ * Extract invoice from markdown using smartagent DualAgentOrchestrator
+ * Validates JSON and retries if invalid
+ */
+async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Promise<IInvoice | null> {
+  const startTime = Date.now();
+  const maxRetries = 2;
+
+  console.log(`      [${queryId}] Invoice: ${markdown.length} chars`);
+
+  // Build the extraction task with document context
+  const taskPrompt = `Extract the invoice data from this document and output ONLY the JSON:
+
+${markdown}
+
+${JSON_EXTRACTION_PROMPT}`;
+
+  try {
+    let result = await orchestrator.run(taskPrompt);
+    let elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] Status: ${result.status}, Iterations: ${result.iterations} (${elapsed}s)`);
+
+    // Try to parse JSON from result
+    let jsonData: Record<string, unknown> | null = null;
+    let responseText = result.result || '';
+
+    if (result.success && responseText) {
+      jsonData = tryExtractJson(responseText);
+    }
+
+    // Fallback: try parsing from history
+    if (!jsonData && result.history?.length > 0) {
+      const lastMessage = result.history[result.history.length - 1];
+      if (lastMessage?.content) {
+        responseText = lastMessage.content;
+        jsonData = tryExtractJson(responseText);
+      }
+    }
+
+    // If JSON is invalid, retry with correction request
+    let retries = 0;
+    while (!jsonData && retries < maxRetries) {
+      retries++;
+      console.log(`      [${queryId}] Invalid JSON, requesting correction (retry ${retries}/${maxRetries})...`);
+
+      result = await orchestrator.continueTask(
+        `Your response was not valid JSON. Please output ONLY the JSON object with no markdown, no explanation, no thinking tags. Just the raw JSON starting with { and ending with }. Format:
+{"invoice_number":"X","invoice_date":"YYYY-MM-DD","vendor_name":"X","currency":"EUR","net_amount":0,"vat_amount":0,"total_amount":0}`
+      );
+
+      elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+      console.log(`      [${queryId}] Retry ${retries}: ${result.status} (${elapsed}s)`);
+
+      responseText = result.result || '';
+      if (responseText) {
+        jsonData = tryExtractJson(responseText);
+      }
+
+      if (!jsonData && result.history?.length > 0) {
+        const lastMessage = result.history[result.history.length - 1];
+        if (lastMessage?.content) {
+          responseText = lastMessage.content;
+          jsonData = tryExtractJson(responseText);
+        }
+      }
+    }
+
+    if (!jsonData) {
+      console.log(`      [${queryId}] Failed to get valid JSON after ${retries} retries`);
+      return null;
+    }
+
+    console.log(`      [${queryId}] Valid JSON extracted`);
+    return {
+      invoice_number: extractInvoiceNumber(String(jsonData.invoice_number || '')),
+      invoice_date: extractDate(String(jsonData.invoice_date || '')),
+      vendor_name: String(jsonData.vendor_name || '').replace(/\*\*/g, '').replace(/`/g, '').trim(),
+      currency: extractCurrency(String(jsonData.currency || '')),
+      net_amount: parseAmount(jsonData.net_amount as string | number),
+      vat_amount: parseAmount(jsonData.vat_amount as string | number),
+      total_amount: parseAmount(jsonData.total_amount as string | number),
+    };
+  } catch (error) {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${queryId}] ERROR: ${error} (${elapsed}s)`);
+    throw error;
+  }
+}
+
+/**
+ * Extract invoice (single pass - GPT-OSS is more reliable)
+ */
+async function extractInvoice(markdown: string, docName: string): Promise<IInvoice> {
+  console.log(`    [${docName}] Extracting...`);
+  const invoice = await extractInvoiceFromMarkdown(markdown, docName);
+  if (!invoice) {
+    return {
+      invoice_number: '',
+      invoice_date: '',
+      vendor_name: '',
+      currency: 'EUR',
+      net_amount: 0,
+      vat_amount: 0,
+      total_amount: 0,
+    };
+  }
+  console.log(`    [${docName}] Extracted: ${invoice.invoice_number}`);
+  return invoice;
+}
+
+/**
+ * Normalize date to YYYY-MM-DD
+ */
+function normalizeDate(dateStr: string | null): string {
+  if (!dateStr) return '';
+  if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
+
+  const monthMap: Record<string, string> = {
+    JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
+    JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
+  };
+
+  let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
+  if (match) {
+    return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
+  }
+
+  match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
+  if (match) {
+    return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
+  }
+
+  return dateStr;
+}
+
+/**
+ * Compare extracted invoice against expected
+ */
+function compareInvoice(
+  extracted: IInvoice,
+  expected: IInvoice
+): { match: boolean; errors: string[] } {
+  const errors: string[] = [];
+
+  const extNum = extracted.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
+  const expNum = expected.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
+  if (extNum !== expNum) {
+    errors.push(`invoice_number: exp "${expected.invoice_number}", got "${extracted.invoice_number}"`);
+  }
+
+  if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
+    errors.push(`invoice_date: exp "${expected.invoice_date}", got "${extracted.invoice_date}"`);
+  }
+
+  if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
+    errors.push(`total_amount: exp ${expected.total_amount}, got ${extracted.total_amount}`);
+  }
+
+  if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
+    errors.push(`currency: exp "${expected.currency}", got "${extracted.currency}"`);
+  }
+
+  return { match: errors.length === 0, errors };
+}
+
+/**
+ * Find all test cases
+ */
+function findTestCases(): ITestCase[] {
+  const testDir = path.join(process.cwd(), '.nogit/invoices');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: ITestCase[] = [];
+
+  for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${baseName}.json`;
+    if (files.includes(jsonFile)) {
+      testCases.push({
+        name: baseName,
+        pdfPath: path.join(testDir, pdf),
+        jsonPath: path.join(testDir, jsonFile),
+      });
+    }
+  }
+
+  return testCases.sort((a, b) => a.name.localeCompare(b.name));
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} invoice test cases\n`);
+
+// Ensure cache directory exists
+if (!fs.existsSync(MD_CACHE_DIR)) {
+  fs.mkdirSync(MD_CACHE_DIR, { recursive: true });
+}
+
+// -------- STAGE 1: OCR with Nanonets --------
+
+tap.test('Stage 1: Convert invoices to markdown (with caching)', async () => {
+  console.log('\n========== STAGE 1: Nanonets OCR ==========\n');
+
+  // Check which invoices need OCR conversion
+  const needsConversion: ITestCase[] = [];
+  let cachedCount = 0;
+
+  for (const tc of testCases) {
+    const mdPath = path.join(MD_CACHE_DIR, `${tc.name}.md`);
+    if (fs.existsSync(mdPath)) {
+      cachedCount++;
+      tc.markdownPath = mdPath;
+      console.log(`  [CACHED] ${tc.name} - using cached markdown`);
+    } else {
+      needsConversion.push(tc);
+    }
+  }
+
+  console.log(`\n  Summary: ${cachedCount} cached, ${needsConversion.length} need conversion\n`);
+
+  if (needsConversion.length === 0) {
+    console.log('  All invoices already cached, skipping Nanonets OCR\n');
+    return;
+  }
+
+  // Start Nanonets only if there are files to convert
+  console.log('  Starting Nanonets for OCR conversion...\n');
+  const ok = await ensureNanonetsOcr();
+  expect(ok).toBeTrue();
+
+  // Convert only the invoices that need conversion
+  for (const tc of needsConversion) {
+    console.log(`\n    === ${tc.name} ===`);
+
+    const images = convertPdfToImages(tc.pdfPath);
+    console.log(`      Pages: ${images.length}`);
+
+    const markdown = await convertDocumentToMarkdown(images, tc.name);
+
+    const mdPath = path.join(MD_CACHE_DIR, `${tc.name}.md`);
+    fs.writeFileSync(mdPath, markdown);
+    tc.markdownPath = mdPath;
+    console.log(`      Saved: ${mdPath}`);
+  }
+
+  console.log(`\n  Stage 1 complete: ${needsConversion.length} invoices converted to markdown\n`);
+});
+
+tap.test('Stage 1: Stop Nanonets', async () => {
+  stopNanonets();
+  await new Promise(resolve => setTimeout(resolve, 3000));
+  expect(isContainerRunning('nanonets-test')).toBeFalse();
+});
+
+// -------- STAGE 2: Extraction with GPT-OSS 20B --------
+
+tap.test('Stage 2: Setup Ollama + GPT-OSS 20B', async () => {
+  console.log('\n========== STAGE 2: GPT-OSS 20B Extraction ==========\n');
+
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
+
+  // Initialize SmartAi and DualAgentOrchestrator
+  console.log('  [SmartAgent] Starting SmartAi...');
+  await smartAi.start();
+
+  console.log('  [SmartAgent] Creating DualAgentOrchestrator with native tool calling...');
+  orchestrator = new DualAgentOrchestrator({
+    smartAiInstance: smartAi,
+    defaultProvider: 'ollama',
+    guardianPolicyPrompt: `
+      JSON EXTRACTION POLICY:
+      - APPROVE all JSON extraction tasks
+      - APPROVE all json.validate tool calls
+      - This is a read-only operation - no file system or network access needed
+      - The task is to extract structured data from document text
+    `,
+    driverSystemMessage: `You are a precise JSON extraction assistant. Your only job is to extract invoice data from documents.
+
+CRITICAL RULES:
+1. Output valid JSON with the exact format requested
+2. If you cannot find a value, use empty string "" or 0 for numbers
+3. Before completing, validate your JSON using the json_validate tool
+4. Only complete after validation passes`,
+    maxIterations: 5,
+    // Enable native tool calling for GPT-OSS (uses Harmony format instead of XML)
+    useNativeToolCalling: true,
+    // Enable streaming for real-time progress visibility
+    onToken: (token, source) => {
+      if (source === 'driver') {
+        process.stdout.write(token);
+      }
+    },
+    // Enable progress events to see tool calls
+    onProgress: (event: { logMessage: string }) => {
+      console.log(event.logMessage);
+    },
+  });
+
+  // Register JsonValidatorTool for self-validation
+  orchestrator.registerTool(new JsonValidatorTool());
+
+  console.log('  [SmartAgent] Starting orchestrator...');
+  await orchestrator.start();
+  console.log('  [SmartAgent] Ready for extraction');
+});
+
+let passedCount = 0;
+let failedCount = 0;
+const processingTimes: number[] = [];
+
+for (const tc of testCases) {
+  tap.test(`Stage 2: Extract ${tc.name}`, async () => {
+    const expected: IInvoice = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    console.log(`\n    === ${tc.name} ===`);
+    console.log(`    Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
+
+    const startTime = Date.now();
+
+    const mdPath = path.join(MD_CACHE_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      throw new Error(`Markdown not found: ${mdPath}. Run Stage 1 first.`);
+    }
+    const markdown = fs.readFileSync(mdPath, 'utf-8');
+    console.log(`    Markdown: ${markdown.length} chars`);
+
+    const extracted = await extractInvoice(markdown, tc.name);
+
+    const elapsedMs = Date.now() - startTime;
+    processingTimes.push(elapsedMs);
+
+    console.log(`    Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
+
+    const result = compareInvoice(extracted, expected);
+
+    if (result.match) {
+      passedCount++;
+      console.log(`    Result: MATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+    } else {
+      failedCount++;
+      console.log(`    Result: MISMATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+      result.errors.forEach(e => console.log(`      - ${e}`));
+    }
+
+    expect(result.match).toBeTrue();
+  });
+}
+
+tap.test('Summary', async () => {
+  // Cleanup orchestrator and SmartAi
+  if (orchestrator) {
+    console.log('\n  [SmartAgent] Stopping orchestrator...');
+    await orchestrator.stop();
+  }
+  console.log('  [SmartAgent] Stopping SmartAi...');
+  await smartAi.stop();
+
+  const totalInvoices = testCases.length;
+  const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
+  const totalTimeMs = processingTimes.reduce((a, b) => a + b, 0);
+  const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
+
+  console.log(`\n========================================`);
+  console.log(`   Invoice Summary (Nanonets + GPT-OSS 20B)`);
+  console.log(`========================================`);
+  console.log(`  Stage 1:   Nanonets-OCR-s (doc -> md)`);
+  console.log(`  Stage 2:   GPT-OSS 20B + SmartAgent (md -> JSON)`);
+  console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
+  console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
+  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
+  console.log(`----------------------------------------`);
+  console.log(`  Total time:   ${(totalTimeMs / 1000).toFixed(1)}s`);
+  console.log(`  Avg per inv:  ${avgTimeSec.toFixed(1)}s`);
+  console.log(`========================================\n`);
+  console.log(`  Cache location: ${MD_CACHE_DIR}\n`);
+});
+
+export default tap.start();