v1.14.0

feat(docker-images): add vLLM-based Nanonets-OCR2-3B image, Qwen3-VL Ollama image and refactor build/docs/tests to use new runtime/layout
update
2026-01-19 21:05:51 +00:00 · 2026-01-19 21:05:51 +00:00 · 2026-01-19 11:51:23 +00:00
14 changed files with 1945 additions and 1238 deletions
--- a/0
+++ b/0
--- a/21
+++ b/21
@@ -1,21 +1,22 @@
-# Nanonets-OCR-s Vision Language Model
-# Based on Qwen2.5-VL-3B, fine-tuned for document OCR
-# ~8-10GB VRAM, outputs structured markdown with semantic tags
+# 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_ocr -t nanonets-ocr .
+# 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-OCR-s - Document OCR optimized Vision Language Model"
+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-OCR-s"
+ENV MODEL_NAME="nanonets/Nanonets-OCR2-3B"
 ENV HOST="0.0.0.0"
 ENV PORT="8000"
-ENV MAX_MODEL_LEN="8192"
+ENV MAX_MODEL_LEN="30000"
 ENV GPU_MEMORY_UTILIZATION="0.9"

 # Expose OpenAI-compatible API port
@@ -25,9 +26,9 @@ EXPOSE 8000
 HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=5 \
    CMD curl -f http://localhost:8000/health || exit 1

-# Start vLLM server with Nanonets-OCR-s model
-CMD ["--model", "nanonets/Nanonets-OCR-s", \
+# Start vLLM server with Nanonets-OCR2-3B model
+CMD ["--model", "nanonets/Nanonets-OCR2-3B", \
     "--trust-remote-code", \
-     "--max-model-len", "8192", \
+     "--max-model-len", "30000", \
     "--host", "0.0.0.0", \
     "--port", "8000"]
--- a/0
+++ b/0
--- a/build-images.sh
+++ b/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"
--- a/changelog.md
+++ b/changelog.md
@@ -1,5 +1,16 @@
 # Changelog

+## 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

--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
  "name": "@host.today/ht-docker-ai",
-  "version": "1.13.2",
+  "version": "1.14.0",
  "type": "module",
  "private": false,
  "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",
@@ -13,8 +13,8 @@
    "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"
  },
  "repository": {
    "type": "git",
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
--- a/readme.hints.md
+++ b/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 |
 |----------|--------|-------------|
@@ -56,192 +75,23 @@ 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)
-
-### 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.
-
-**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
-
-### Docker Images
-
-| Tag | Description |
-|-----|-------------|
-| `paddleocr-vl` | GPU variant using vLLM (recommended) |
-| `paddleocr-vl-cpu` | CPU variant using transformers |
-
-### API Endpoints (OpenAI-compatible)
-
-| Endpoint | Method | Description |
-|----------|--------|-------------|
-| `/health` | GET | Health check with model info |
-| `/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",
-  "messages": [
-    {
-      "role": "user",
-      "content": [
-        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
-        {"type": "text", "text": "Table Recognition:"}
-      ]
-    }
-  ],
-  "temperature": 0.0,
-  "max_tokens": 8192
-}
-```
-
-**Task Prompts:**
- `"OCR:"` - Text recognition
- `"Table Recognition:"` - Table extraction (returns markdown)
- `"Formula Recognition:"` - Formula extraction
- `"Chart Recognition:"` - Chart extraction
-
-**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"
-    }
-  ]
-}
-```
-
-### 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
-
---
-
-## Adding New Models
-
-To add a new model variant:
-
-1. Create `Dockerfile_<modelname>`
-2. Set `MODEL_NAME` environment variable
-3. Update `build-images.sh` with new build target
-4. Add documentation to `readme.md`
-
-## Troubleshooting
-
-### Model download hangs
-
-Check container logs:
-```bash
-docker logs -f <container-name>
-```
-
-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`
-
-### API not responding
-
-1. Check if container is healthy: `docker ps`
-2. Check logs for errors: `docker logs <container>`
-3. Verify port mapping: `curl localhost:11434/api/tags`
-
-## CI/CD Integration
-
-Build and push using npmci:
-
-```bash
-npmci docker login
-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.
-
 ---

 ## Nanonets-OCR-s
@@ -254,7 +104,7 @@ Nanonets-OCR-s is a Qwen2.5-VL-3B model fine-tuned specifically for document OCR
 - Based on Qwen2.5-VL-3B (~4B parameters)
 - Fine-tuned for document OCR
 - Outputs markdown with semantic HTML tags
- ~8-10GB VRAM (fits comfortably in 16GB)
+- ~10GB VRAM

 ### Docker Images

@@ -305,7 +155,7 @@ Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.
 ### Performance

 - **GPU (vLLM)**: ~3-8 seconds per page
- **VRAM usage**: ~8-10GB
+- **VRAM usage**: ~10GB

 ### Two-Stage Pipeline (Nanonets + Qwen3)

@@ -332,6 +182,76 @@ docker start minicpm-test

 ---

+## 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
+
+---
+
+## Adding New Models
+
+To add a new model variant:
+
+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`
+
+## Troubleshooting
+
+### Model download hangs
+
+Check container logs:
+```bash
+docker logs -f <container-name>
+```
+
+The model download is ~5GB and may take several minutes.
+
+### Out of memory
+
+- GPU: Use a lighter model variant or upgrade VRAM
+- Add more GPU memory: Consider multi-GPU setup
+
+### API not responding
+
+1. Check if container is healthy: `docker ps`
+2. Check logs for errors: `docker logs <container>`
+3. Verify port mapping: `curl localhost:11434/api/tags`
+
+## CI/CD Integration
+
+Build and push using npmci:
+
+```bash
+npmci docker login
+npmci docker build
+npmci docker push code.foss.global
+```
+
+---
+
 ## Related Resources

 - [Ollama Documentation](https://ollama.ai/docs)
--- a/readme.md
+++ b/readme.md
@@ -1,40 +1,45 @@
 # @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 lightweight 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-OCR-s** | ~4B | Document OCR with semantic markdown output | OpenAI-compatible | 8000 | ~10GB |
+| **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-OCR-s | vLLM | 8000 | ~10GB |
+| `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**.

 ### Quick Start

-**GPU (Recommended):**
 ```bash
 docker run -d \
  --name minicpm \
@@ -44,15 +49,6 @@ 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
@@ -85,108 +81,128 @@ 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-OCR-s

-A specialized 0.9B Vision-Language Model optimized for document parsing. Native support for tables, formulas, charts, and text extraction in 109 languages.
+A **Qwen2.5-VL-3B** model fine-tuned specifically for document OCR. Outputs structured markdown with semantic HTML tags—perfect for preserving document structure.
+
+### Key Features
+
+- 📝 **Semantic output:** Tables → HTML, equations → LaTeX, watermarks/page numbers → tagged
+- 🌍 **Multilingual:** Inherits Qwen's broad language support
+- ⚡ **Efficient:** ~10GB VRAM, runs great on consumer GPUs
+- 🔌 **OpenAI-compatible:** Drop-in replacement for existing pipelines

 ### 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-OCR-s",
    "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-OCR-s returns markdown with semantic tags:

-### API Endpoints
-
-| 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 |
-
-### Image Input Methods
-
-PaddleOCR-VL accepts images in multiple formats:
-
-```javascript
-// Base64 data URL
-"data:image/png;base64,iVBORw0KGgo..."
-
-// HTTP URL
-"https://example.com/document.png"
-
-// Raw base64
-"iVBORw0KGgo..."
-```
-
-**Supported formats:** PNG, JPEG, WebP, BMP, GIF, TIFF
-
-**Optimal resolution:** 1080p–2K. Images are automatically scaled for best results.
+| 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>` |

 ### Performance

-| Mode | Speed per Page |
-|------|----------------|
-| GPU (CUDA) | 2–5 seconds |
-| CPU | 30–60 seconds |
+| Metric | Value |
+|--------|-------|
+| Speed | 3–8 seconds per page |
+| VRAM | ~10GB |
+
+---
+
+## 🧠 Qwen3-VL-30B-A3B
+
+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.
+
+### 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
+```
+
+Then pull the model (one-time, ~20GB):
+```bash
+docker exec qwen3vl ollama pull qwen3-vl:30b-a3b
+```
+
+### API Usage
+
+```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>"]
+  }]
+}'
+```
+
+### 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:
@@ -206,9 +222,9 @@ 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 +247,7 @@ volumes:

 ## ⚙️ Environment Variables

-### MiniCPM-V 4.5
+### MiniCPM-V 4.5 & Qwen3-VL (Ollama-based)

 | Variable | Default | Description |
 |----------|---------|-------------|
@@ -239,13 +255,46 @@ 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-OCR-s` | HuggingFace model ID |
+| `HOST` | `0.0.0.0` | API bind address |
+| `PORT` | `8000` | API port |
+| `MAX_MODEL_LEN` | `8192` | 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-OCR-s 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-OCR-s |
+| Complex visual reasoning / code generation | Qwen3-VL-30B |
+| Multi-image analysis | MiniCPM-V 4.5 |
+| Visual agent tasks | Qwen3-VL-30B |

 ---

@@ -265,37 +314,16 @@ cd ht-docker-ai

 ---

-## 🏗️ 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 +343,13 @@ 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
--- a/test/helpers/docker.ts
+++ b/test/helpers/docker.ts
@@ -32,10 +32,10 @@ export const IMAGES = {
    healthTimeout: 120000,
  } as IImageConfig,

-  // Nanonets-OCR-s - Document OCR optimized VLM (Qwen2.5-VL-3B fine-tuned)
+  // Nanonets-OCR2-3B - Document OCR optimized VLM (Qwen2.5-VL-3B fine-tuned, Oct 2025)
  nanonetsOcr: {
    name: 'nanonets-ocr',
-    dockerfile: 'Dockerfile_nanonets_ocr',
+    dockerfile: 'Dockerfile_nanonets_vllm_gpu_VRAM10GB',
    buildContext: '.',
    containerName: 'nanonets-test',
    ports: ['8000:8000'],
@@ -340,12 +340,12 @@ export async function ensureQwen3Vl(): Promise<boolean> {
 }

 /**
- * Ensure Nanonets-OCR-s service is running (via vLLM)
- * Document OCR optimized VLM based on Qwen2.5-VL-3B
+ * 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-OCR-s requires GPU, but none detected');
+    console.log('[Docker] WARNING: Nanonets-OCR2-3B requires GPU, but none detected');
  }
  return ensureService(IMAGES.nanonetsOcr);
 }
--- a/test/test.bankstatements.nanonets.ts
+++ b/test/test.bankstatements.nanonets.ts
@@ -1,7 +1,7 @@
 /**
- * Bank statement extraction using Nanonets-OCR-s + GPT-OSS 20B (sequential two-stage pipeline)
+ * Bank statement extraction using Nanonets-OCR2-3B + GPT-OSS 20B (sequential two-stage pipeline)
 *
- * Stage 1: Nanonets-OCR-s converts ALL document pages to markdown (stop after completion)
+ * 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.
@@ -14,7 +14,7 @@ import * as os from 'os';
 import { ensureNanonetsOcr, ensureMiniCpm, removeContainer, isContainerRunning } from './helpers/docker.js';

 const NANONETS_URL = 'http://localhost:8000/v1';
-const NANONETS_MODEL = 'nanonets/Nanonets-OCR-s';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR2-3B';

 const OLLAMA_URL = 'http://localhost:11434';
 const EXTRACTION_MODEL = 'gpt-oss:20b';
@@ -28,12 +28,19 @@ interface ITransaction {
  amount: number;
 }

+interface IImageData {
+  base64: string;
+  width: number;
+  height: number;
+  pageNum: number;
+}
+
 interface ITestCase {
  name: string;
  pdfPath: string;
  jsonPath: string;
  markdownPath?: string;
-  images?: string[];
+  images?: IImageData[];
 }

 // Nanonets-specific prompt for document OCR to markdown
@@ -50,12 +57,31 @@ const JSON_EXTRACTION_PROMPT = `Extract ALL transactions from this bank statemen
 STATEMENT:
 `;

+// 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
+
 /**
- * Convert PDF to PNG images using ImageMagick
+ * Estimate visual tokens for an image based on dimensions
 */
-function convertPdfToImages(pdfPath: string): string[] {
+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.png');
+  const outputPattern = path.join(tempDir, 'page-%d.jpg');

  try {
    execSync(
@@ -63,13 +89,24 @@ function convertPdfToImages(pdfPath: string): string[] {
      { stdio: 'pipe' }
    );

-    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.png')).sort();
-    const images: string[] = [];
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
+    const images: IImageData[] = [];

-    for (const file of files) {
+    for (let i = 0; i < files.length; i++) {
+      const file = files[i];
      const imagePath = path.join(tempDir, file);
      const imageData = fs.readFileSync(imagePath);
-      images.push(imageData.toString('base64'));
+
+      // 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;
@@ -79,10 +116,28 @@ function convertPdfToImages(pdfPath: string): string[] {
 }

 /**
- * Convert a single page to markdown using Nanonets-OCR-s
+ * Convert a batch of pages to markdown using Nanonets-OCR-s
 */
-async function convertPageToMarkdown(image: string, pageNum: number): Promise<string> {
+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(`${NANONETS_URL}/chat/completions`, {
    method: 'POST',
@@ -94,14 +149,12 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
      model: NANONETS_MODEL,
      messages: [{
        role: 'user',
-        content: [
-          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
-          { type: 'text', text: NANONETS_OCR_PROMPT },
-        ],
+        content,
      }],
-      max_tokens: 4096,
+      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);
@@ -112,25 +165,35 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
  }

  const data = await response.json();
-  const content = (data.choices?.[0]?.message?.content || '').trim();
-  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
-  return content;
+  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
+ * Convert all pages of a document to markdown using smart batching
 */
-async function convertDocumentToMarkdown(images: string[], docName: string): Promise<string> {
-  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
+  const batches = batchImages(images);
+  console.log(`      [${docName}] Processing ${images.length} page(s) in ${batches.length} batch(es)...`);

-  const markdownPages: string[] = [];
+  const markdownParts: string[] = [];

-  for (let i = 0; i < images.length; i++) {
-    const markdown = await convertPageToMarkdown(images[i], i + 1);
-    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+  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 = markdownPages.join('\n\n');
+  const fullMarkdown = markdownParts.join('\n\n');
  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
  return fullMarkdown;
 }
@@ -161,25 +224,6 @@ async function ensureExtractionModel(): Promise<boolean> {
      const models = data.models || [];
      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
-
-        // Warmup: send a simple request to ensure model is loaded
-        console.log(`    [Ollama] Warming up model...`);
-        const warmupResponse = await fetch(`${OLLAMA_URL}/api/chat`, {
-          method: 'POST',
-          headers: { 'Content-Type': 'application/json' },
-          body: JSON.stringify({
-            model: EXTRACTION_MODEL,
-            messages: [{ role: 'user', content: 'Return: [{"test": 1}]' }],
-            stream: false,
-          }),
-          signal: AbortSignal.timeout(120000),
-        });
-
-        if (warmupResponse.ok) {
-          const warmupData = await warmupResponse.json();
-          console.log(`    [Ollama] Warmup complete (${warmupData.message?.content?.length || 0} chars)`);
-        }
-
        return true;
      }
    }
@@ -201,22 +245,24 @@ async function ensureExtractionModel(): Promise<boolean> {
 * Extract transactions from markdown using GPT-OSS 20B (streaming)
 */
 async function extractTransactionsFromMarkdown(markdown: string, queryId: string): Promise<ITransaction[]> {
-  console.log(`      [${queryId}] Sending to ${EXTRACTION_MODEL}...`);
-  console.log(`      [${queryId}] Markdown length: ${markdown.length}`);
  const startTime = Date.now();
-
  const fullPrompt = JSON_EXTRACTION_PROMPT + markdown;
-  console.log(`      [${queryId}] Prompt preview: ${fullPrompt.substring(0, 200)}...`);
+
+  // Log exact prompt
+  console.log(`\n      [${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: fullPrompt,
-      }],
+      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),  // 10 minute timeout
@@ -228,24 +274,45 @@ async function extractTransactionsFromMarkdown(markdown: string, queryId: string
    throw new Error(`Ollama API error: ${response.status}`);
  }

-  // Stream the response and log to console
+  // Stream the response
  let content = '';
+  let thinkingContent = '';
+  let thinkingStarted = false;
+  let outputStarted = false;
  const reader = response.body!.getReader();
  const decoder = new TextDecoder();

-  process.stdout.write(`      [${queryId}] `);
-
+  try {
    while (true) {
      const { done, value } = await reader.read();
      if (done) break;

      const chunk = decoder.decode(value, { stream: true });
+
      // Each line is a JSON object
      for (const line of chunk.split('\n').filter(l => l.trim())) {
        try {
          const json = JSON.parse(line);
+
+          // Stream thinking tokens
+          const thinking = json.message?.thinking || '';
+          if (thinking) {
+            if (!thinkingStarted) {
+              process.stdout.write(`      [${queryId}] THINKING: `);
+              thinkingStarted = true;
+            }
+            process.stdout.write(thinking);
+            thinkingContent += thinking;
+          }
+
+          // Stream content tokens
          const token = json.message?.content || '';
          if (token) {
+            if (!outputStarted) {
+              if (thinkingStarted) process.stdout.write('\n');
+              process.stdout.write(`      [${queryId}] OUTPUT: `);
+              outputStarted = true;
+            }
            process.stdout.write(token);
            content += token;
          }
@@ -254,9 +321,12 @@ async function extractTransactionsFromMarkdown(markdown: string, queryId: string
        }
      }
    }
+  } finally {
+    if (thinkingStarted || outputStarted) process.stdout.write('\n');
+  }

  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
-  console.log(`\n      [${queryId}] Done: ${content.length} chars (${elapsed}s)`);
+  console.log(`      [${queryId}] Done: ${thinkingContent.length} thinking chars, ${content.length} output chars (${elapsed}s)`);

  return parseJsonResponse(content, queryId);
 }
--- a/test/test.invoices.extraction.ts
+++ b/test/test.invoices.extraction.ts
@@ -0,0 +1,436 @@
+/**
+ * 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${markdown}` },
+        { role: 'assistant', content: 'I have read the invoice document you provided. I can see all the text content. What would you like me to do with it?' },
+        { role: 'user', content: JSON_EXTRACTION_PROMPT },
+      ],
+      stream: true,
+    }),
+    signal: AbortSignal.timeout(120000), // 2 min timeout
+  });
+
+  if (!response.ok) {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+    console.log(`      [${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();
--- a/test/test.invoices.failed.ts
+++ b/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();
--- a/test/test.invoices.nanonets.ts
+++ b/test/test.invoices.nanonets.ts
@@ -1,8 +1,8 @@
 /**
- * Invoice extraction using Nanonets-OCR-s + Qwen3 (sequential two-stage pipeline)
+ * Invoice extraction using Nanonets-OCR2-3B + GPT-OSS 20B (sequential two-stage pipeline)
 *
- * Stage 1: Nanonets-OCR-s converts ALL document pages to markdown (stop after completion)
- * Stage 2: Qwen3 extracts structured JSON from saved markdown (after Nanonets stops)
+ * 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.
 */
@@ -14,10 +14,10 @@ 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-OCR-s';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR2-3B';

 const OLLAMA_URL = 'http://localhost:11434';
-const QWEN_MODEL = 'qwen3:8b';
+const EXTRACTION_MODEL = 'gpt-oss:20b';

 // Temp directory for storing markdown between stages
 const TEMP_MD_DIR = path.join(os.tmpdir(), 'nanonets-invoices-markdown');
@@ -32,6 +32,13 @@ interface IInvoice {
  total_amount: number;
 }

+interface IImageData {
+  base64: string;
+  width: number;
+  height: number;
+  pageNum: number;
+}
+
 interface ITestCase {
  name: string;
  pdfPath: string;
@@ -47,7 +54,7 @@ If there is an image in the document and image caption is not present, add a sma
 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 Qwen3
+// 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:
@@ -73,12 +80,31 @@ Return ONLY this JSON format, no explanation:
 INVOICE TEXT:
 `;

+// 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
+
 /**
- * Convert PDF to PNG images
+ * Estimate visual tokens for an image based on dimensions
 */
-function convertPdfToImages(pdfPath: string): string[] {
+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.png');
+  const outputPattern = path.join(tempDir, 'page-%d.jpg');

  try {
    execSync(
@@ -86,13 +112,24 @@ function convertPdfToImages(pdfPath: string): string[] {
      { stdio: 'pipe' }
    );

-    const files = fs.readdirSync(tempDir).filter((f) => f.endsWith('.png')).sort();
-    const images: string[] = [];
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
+    const images: IImageData[] = [];

-    for (const file of files) {
+    for (let i = 0; i < files.length; i++) {
+      const file = files[i];
      const imagePath = path.join(tempDir, file);
      const imageData = fs.readFileSync(imagePath);
-      images.push(imageData.toString('base64'));
+
+      // 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;
@@ -102,10 +139,28 @@ function convertPdfToImages(pdfPath: string): string[] {
 }

 /**
- * Convert a single page to markdown using Nanonets-OCR-s
+ * Convert a batch of pages to markdown using Nanonets-OCR-s
 */
-async function convertPageToMarkdown(image: string, pageNum: number): Promise<string> {
+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(`${NANONETS_URL}/chat/completions`, {
    method: 'POST',
@@ -117,14 +172,12 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
      model: NANONETS_MODEL,
      messages: [{
        role: 'user',
-        content: [
-          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
-          { type: 'text', text: NANONETS_OCR_PROMPT },
-        ],
+        content,
      }],
-      max_tokens: 4096,
+      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);
@@ -135,25 +188,35 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
  }

  const data = await response.json();
-  const content = (data.choices?.[0]?.message?.content || '').trim();
-  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
-  return content;
+  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
+ * Convert all pages of a document to markdown using smart batching
 */
-async function convertDocumentToMarkdown(images: string[], docName: string): Promise<string> {
-  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
+  const batches = batchImages(images);
+  console.log(`      [${docName}] Processing ${images.length} page(s) in ${batches.length} batch(es)...`);

-  const markdownPages: string[] = [];
+  const markdownParts: string[] = [];

-  for (let i = 0; i < images.length; i++) {
-    const markdown = await convertPageToMarkdown(images[i], i + 1);
-    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+  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 = markdownPages.join('\n\n');
+  const fullMarkdown = markdownParts.join('\n\n');
  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
  return fullMarkdown;
 }
@@ -173,16 +236,16 @@ function stopNanonets(): void {
 }

 /**
- * Ensure Qwen3 model is available
+ * Ensure GPT-OSS 20B model is available
 */
-async function ensureQwen3(): Promise<boolean> {
+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 === QWEN_MODEL)) {
-        console.log(`    [Ollama] Model available: ${QWEN_MODEL}`);
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
        return true;
      }
    }
@@ -190,11 +253,11 @@ async function ensureQwen3(): Promise<boolean> {
    return false;
  }

-  console.log(`    [Ollama] Pulling ${QWEN_MODEL}...`);
+  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: QWEN_MODEL, stream: false }),
+    body: JSON.stringify({ name: EXTRACTION_MODEL, stream: false }),
  });

  return pullResponse.ok;
@@ -303,88 +366,102 @@ function parseJsonToInvoice(response: string): IInvoice | null {
 }

 /**
- * Extract invoice from markdown using Qwen3
+ * Extract invoice from markdown using GPT-OSS 20B (streaming)
 */
 async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Promise<IInvoice | null> {
-  console.log(`      [${queryId}] Sending to ${QWEN_MODEL}...`);
  const startTime = Date.now();
+  const fullPrompt = JSON_EXTRACTION_PROMPT + markdown;
+
+  // Log exact prompt
+  console.log(`\n      [${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' },
-    signal: AbortSignal.timeout(600000), // 10 minute timeout for large documents
    body: JSON.stringify({
-      model: QWEN_MODEL,
-      messages: [{
-        role: 'user',
-        content: JSON_EXTRACTION_PROMPT + markdown,
-      }],
-      stream: false,
-      options: {
-        num_predict: 2000,
-        temperature: 0.1,
-      },
+      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), // 10 minute timeout for large documents
  });

-  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
-
  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}`);
  }

-  const data = await response.json();
-  const content = (data.message?.content || '').trim();
-  console.log(`      [${queryId}] Response: ${content.length} chars (${elapsed}s)`);
+  // 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 });
+
+      // Each line is a JSON object
+      for (const line of chunk.split('\n').filter(l => l.trim())) {
+        try {
+          const json = JSON.parse(line);
+
+          // Stream thinking tokens
+          const thinking = json.message?.thinking || '';
+          if (thinking) {
+            if (!thinkingStarted) {
+              process.stdout.write(`      [${queryId}] THINKING: `);
+              thinkingStarted = true;
+            }
+            process.stdout.write(thinking);
+            thinkingContent += thinking;
+          }
+
+          // Stream content tokens
+          const token = json.message?.content || '';
+          if (token) {
+            if (!outputStarted) {
+              if (thinkingStarted) process.stdout.write('\n');
+              process.stdout.write(`      [${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)`);

  return parseJsonToInvoice(content);
 }

 /**
- * Compare two invoices for consensus
+ * Extract invoice (single pass - GPT-OSS is more reliable)
 */
-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 with consensus
- */
-async function extractWithConsensus(markdown: string, docName: string): Promise<IInvoice> {
-  const MAX_ATTEMPTS = 3;
-
-  for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt++) {
-    console.log(`    [${docName}] Attempt ${attempt}/${MAX_ATTEMPTS}`);
-
-    const inv1 = await extractInvoiceFromMarkdown(markdown, `${docName}-A${attempt}Q1`);
-    const inv2 = await extractInvoiceFromMarkdown(markdown, `${docName}-A${attempt}Q2`);
-
-    if (!inv1 || !inv2) {
-      console.log(`    [${docName}] Parsing failed, retrying...`);
-      continue;
-    }
-
-    console.log(`    [${docName}] Q1: ${inv1.invoice_number} | ${inv1.invoice_date} | ${inv1.total_amount}`);
-    console.log(`    [${docName}] Q2: ${inv2.invoice_number} | ${inv2.invoice_date} | ${inv2.total_amount}`);
-
-    if (invoicesMatch(inv1, inv2)) {
-      console.log(`    [${docName}] CONSENSUS`);
-      return inv2;
-    }
-    console.log(`    [${docName}] No consensus`);
-  }
-
-  // Fallback
-  const fallback = await extractInvoiceFromMarkdown(markdown, `${docName}-FALLBACK`);
-  if (fallback) {
-    console.log(`    [${docName}] FALLBACK: ${fallback.invoice_number} | ${fallback.invoice_date} | ${fallback.total_amount}`);
-    return fallback;
-  }
-
+async function extractInvoice(markdown: string, docName: string): Promise<IInvoice> {
+  console.log(`    [${docName}] Extracting...`);
+  const invoice = await extractInvoiceFromMarkdown(markdown, docName);
+  if (!invoice) {
    return {
      invoice_number: '',
      invoice_date: '',
@@ -394,6 +471,9 @@ async function extractWithConsensus(markdown: string, docName: string): Promise<
      vat_amount: 0,
      total_amount: 0,
    };
+  }
+  console.log(`    [${docName}] Extracted: ${invoice.invoice_number}`);
+  return invoice;
 }

 /**
@@ -520,16 +600,16 @@ tap.test('Stage 1: Stop Nanonets', async () => {
  expect(isContainerRunning('nanonets-test')).toBeFalse();
 });

-// -------- STAGE 2: Extraction with Qwen3 --------
+// -------- STAGE 2: Extraction with GPT-OSS 20B --------

-tap.test('Stage 2: Setup Ollama + Qwen3', async () => {
-  console.log('\n========== STAGE 2: Qwen3 Extraction ==========\n');
+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 qwenOk = await ensureQwen3();
-  expect(qwenOk).toBeTrue();
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
 });

 let passedCount = 0;
@@ -551,7 +631,7 @@ for (const tc of testCases) {
    const markdown = fs.readFileSync(mdPath, 'utf-8');
    console.log(`    Markdown: ${markdown.length} chars`);

-    const extracted = await extractWithConsensus(markdown, tc.name);
+    const extracted = await extractInvoice(markdown, tc.name);

    const elapsedMs = Date.now() - startTime;
    processingTimes.push(elapsedMs);
@@ -580,10 +660,10 @@ tap.test('Summary', async () => {
  const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;

  console.log(`\n========================================`);
-  console.log(`   Invoice Summary (Nanonets + Qwen3)`);
+  console.log(`   Invoice Summary (Nanonets + GPT-OSS 20B)`);
  console.log(`========================================`);
  console.log(`  Stage 1:   Nanonets-OCR-s (doc -> md)`);
-  console.log(`  Stage 2:   Qwen3 8B (md -> JSON)`);
+  console.log(`  Stage 2:   GPT-OSS 20B (md -> JSON)`);
  console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
  console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
Author	SHA1	Message	Date
Juergen Kunz	235aa1352b	v1.14.0 Some checks failed Docker (tags) / security (push) Successful in 29s Details Docker (tags) / test (push) Failing after 39s Details Docker (tags) / release (push) Has been skipped Details Docker (tags) / metadata (push) Has been skipped Details	2026-01-19 21:05:51 +00:00
Juergen Kunz	08728ada4d	feat(docker-images): add vLLM-based Nanonets-OCR2-3B image, Qwen3-VL Ollama image and refactor build/docs/tests to use new runtime/layout	2026-01-19 21:05:51 +00:00
Juergen Kunz	b58bcabc76	update	2026-01-19 11:51:23 +00:00