v1.13.2

Dockerfile_nanonets_ocr

@@ -0,0 +1,33 @@
+# 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
+#
+# Build: docker build -f Dockerfile_nanonets_ocr -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 org.opencontainers.image.source="https://code.foss.global/host.today/ht-docker-ai"
+
+# Environment configuration
+ENV MODEL_NAME="nanonets/Nanonets-OCR-s"
+ENV HOST="0.0.0.0"
+ENV PORT="8000"
+ENV MAX_MODEL_LEN="8192"
+ENV GPU_MEMORY_UTILIZATION="0.9"
+
+# Expose OpenAI-compatible API port
+EXPOSE 8000
+
+# Health check - vLLM exposes /health endpoint
+HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=5 \
+    CMD curl -f http://localhost:8000/health || exit 1
+
+# Start vLLM server with Nanonets-OCR-s model
+CMD ["--model", "nanonets/Nanonets-OCR-s", \
+     "--trust-remote-code", \
+     "--max-model-len", "8192", \
+     "--host", "0.0.0.0", \
+     "--port", "8000"]

changelog.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## 2026-01-18 - 1.13.2 - fix(tests)
+stabilize OCR extraction tests and manage GPU containers
+
+- Add stopAllGpuContainers() and call it before starting GPU images to free GPU memory.
+- Remove PaddleOCR-VL image configs and associated ensure helpers from docker test helper to simplify images list.
+- Split invoice/bankstatement tests into two sequential stages: Stage 1 runs Nanonets OCR to produce markdown files, Stage 2 stops Nanonets and runs model extraction from saved markdown (avoids GPU contention).
+- Introduce temporary markdown directory handling and cleanup; add stopNanonets() and container running checks in tests.
+- Switch bank statement extraction model from qwen3:8b to gpt-oss:20b; add request timeout and improved logging/console output across tests.
+- Refactor extractWithConsensus and extraction functions to accept document identifiers, improve error messages and JSON extraction robustness.
+
 ## 2026-01-18 - 1.13.1 - fix(image_support_files)
 remove PaddleOCR-VL server scripts from image_support_files
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@host.today/ht-docker-ai",
-  "version": "1.13.1",
+  "version": "1.13.2",
   "type": "module",
   "private": false,
   "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",

readme.hints.md

@@ -244,8 +244,97 @@ The bank statement extraction uses a dual-VLM consensus approach:
 
 ---
 
+## Nanonets-OCR-s
+
+### Overview
+
+Nanonets-OCR-s is a Qwen2.5-VL-3B model fine-tuned specifically for document OCR tasks. It outputs structured markdown with semantic tags.
+
+**Key features:**
+- 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)
+
+### Docker Images
+
+| Tag | Description |
+|-----|-------------|
+| `nanonets-ocr` | GPU variant using vLLM (OpenAI-compatible API) |
+
+### API Endpoints (OpenAI-compatible via vLLM)
+
+| Endpoint | Method | Description |
+|----------|--------|-------------|
+| `/health` | GET | Health check |
+| `/v1/models` | GET | List available models |
+| `/v1/chat/completions` | POST | OpenAI-compatible chat completions |
+
+### Request/Response Format
+
+**POST /v1/chat/completions (OpenAI-compatible)**
+```json
+{
+  "model": "nanonets/Nanonets-OCR-s",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
+        {"type": "text", "text": "Extract the text from the above document..."}
+      ]
+    }
+  ],
+  "temperature": 0.0,
+  "max_tokens": 4096
+}
+```
+
+### Nanonets OCR Prompt
+
+The model is designed to work with a specific prompt format:
+```
+Extract the text from the above document as if you were reading it naturally.
+Return the tables in html format.
+Return the equations in LaTeX representation.
+If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
+Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
+Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.
+```
+
+### Performance
+
+- **GPU (vLLM)**: ~3-8 seconds per page
+- **VRAM usage**: ~8-10GB
+
+### Two-Stage Pipeline (Nanonets + Qwen3)
+
+The Nanonets tests use a two-stage pipeline:
+1. **Stage 1**: Nanonets-OCR-s converts images to markdown (via vLLM on port 8000)
+2. **Stage 2**: Qwen3 8B extracts structured JSON from markdown (via Ollama on port 11434)
+
+**GPU Limitation**: Both vLLM and Ollama require significant GPU memory. On a single GPU system:
+- Running both simultaneously causes memory contention
+- For single GPU: Run services sequentially (stop Nanonets before Qwen3)
+- For multi-GPU: Assign each service to a different GPU
+
+**Sequential Execution**:
+```bash
+# Step 1: Run Nanonets OCR (converts to markdown)
+docker start nanonets-test
+# ... perform OCR ...
+docker stop nanonets-test
+
+# Step 2: Run Qwen3 extraction (from markdown)
+docker start minicpm-test
+# ... extract JSON ...
+```
+
+---
+
 ## Related Resources
 
 - [Ollama Documentation](https://ollama.ai/docs)
 - [MiniCPM-V GitHub](https://github.com/OpenBMB/MiniCPM-V)
 - [Ollama API Reference](https://github.com/ollama/ollama/blob/main/docs/api.md)
+- [Nanonets-OCR-s on HuggingFace](https://huggingface.co/nanonets/Nanonets-OCR-s)

test/helpers/docker.ts

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

test/test.bankstatements.nanonets.ts

@@ -0,0 +1,585 @@
+/**
+ * Bank statement extraction using Nanonets-OCR-s + GPT-OSS 20B (sequential two-stage pipeline)
+ *
+ * Stage 1: Nanonets-OCR-s converts ALL document pages to markdown (stop after completion)
+ * Stage 2: GPT-OSS 20B extracts structured JSON from saved markdown (after Nanonets stops)
+ *
+ * This approach avoids GPU contention by running services sequentially.
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { execSync } from 'child_process';
+import * as os from 'os';
+import { ensureNanonetsOcr, ensureMiniCpm, removeContainer, isContainerRunning } from './helpers/docker.js';
+
+const NANONETS_URL = 'http://localhost:8000/v1';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR-s';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const EXTRACTION_MODEL = 'gpt-oss:20b';
+
+// Temp directory for storing markdown between stages
+const TEMP_MD_DIR = path.join(os.tmpdir(), 'nanonets-markdown');
+
+interface ITransaction {
+  date: string;
+  counterparty: string;
+  amount: number;
+}
+
+interface ITestCase {
+  name: string;
+  pdfPath: string;
+  jsonPath: string;
+  markdownPath?: string;
+  images?: 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 = `Extract ALL transactions from this bank statement as JSON array. Each transaction: {"date": "YYYY-MM-DD", "counterparty": "NAME", "amount": -25.99}. Amount negative for debits, positive for credits. Only include actual transactions, not balances. Return ONLY JSON array, no explanation.
+
+STATEMENT:
+`;
+
+/**
+ * Convert PDF to PNG images using ImageMagick
+ */
+function convertPdfToImages(pdfPath: string): string[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
+  const outputPattern = path.join(tempDir, 'page-%d.png');
+
+  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('.png')).sort();
+    const images: string[] = [];
+
+    for (const file of files) {
+      const imagePath = path.join(tempDir, file);
+      const imageData = fs.readFileSync(imagePath);
+      images.push(imageData.toString('base64'));
+    }
+
+    return images;
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Convert a single page to markdown using Nanonets-OCR-s
+ */
+async function convertPageToMarkdown(image: string, pageNum: number): Promise<string> {
+  const startTime = Date.now();
+
+  const response = await fetch(`${NANONETS_URL}/chat/completions`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': 'Bearer dummy',
+    },
+    body: JSON.stringify({
+      model: NANONETS_MODEL,
+      messages: [{
+        role: 'user',
+        content: [
+          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
+          { type: 'text', text: NANONETS_OCR_PROMPT },
+        ],
+      }],
+      max_tokens: 4096,
+      temperature: 0.0,
+    }),
+  });
+
+  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();
+  const content = (data.choices?.[0]?.message?.content || '').trim();
+  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
+  return content;
+}
+
+/**
+ * Convert all pages of a document to markdown
+ */
+async function convertDocumentToMarkdown(images: string[], docName: string): Promise<string> {
+  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+
+  const markdownPages: string[] = [];
+
+  for (let i = 0; i < images.length; i++) {
+    const markdown = await convertPageToMarkdown(images[i], i + 1);
+    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+  }
+
+  const fullMarkdown = markdownPages.join('\n\n');
+  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
+  return fullMarkdown;
+}
+
+/**
+ * Stop Nanonets container
+ */
+function stopNanonets(): void {
+  console.log('    [Docker] Stopping Nanonets container...');
+  try {
+    execSync('docker stop nanonets-test 2>/dev/null || true', { stdio: 'pipe' });
+    // Wait for GPU memory to be released
+    execSync('sleep 5', { stdio: 'pipe' });
+    console.log('    [Docker] Nanonets stopped');
+  } catch {
+    console.log('    [Docker] Nanonets was not running');
+  }
+}
+
+/**
+ * Ensure GPT-OSS 20B model is available and warmed up
+ */
+async function ensureExtractionModel(): Promise<boolean> {
+  try {
+    const response = await fetch(`${OLLAMA_URL}/api/tags`);
+    if (response.ok) {
+      const data = await response.json();
+      const models = data.models || [];
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
+
+        // Warmup: send a simple request to ensure model is loaded
+        console.log(`    [Ollama] Warming up model...`);
+        const warmupResponse = await fetch(`${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;
+      }
+    }
+  } 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;
+}
+
+/**
+ * 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: ${fullPrompt.substring(0, 200)}...`);
+
+  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: EXTRACTION_MODEL,
+      messages: [{
+        role: 'user',
+        content: fullPrompt,
+      }],
+      stream: true,
+    }),
+    signal: AbortSignal.timeout(600000),  // 10 minute 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: ${response.status}`);
+  }
+
+  // Stream the response and log to console
+  let content = '';
+  const reader = response.body!.getReader();
+  const decoder = new TextDecoder();
+
+  process.stdout.write(`      [${queryId}] `);
+
+  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);
+        const token = json.message?.content || '';
+        if (token) {
+          process.stdout.write(token);
+          content += token;
+        }
+      } catch {
+        // Ignore parse errors for partial chunks
+      }
+    }
+  }
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  console.log(`\n      [${queryId}] Done: ${content.length} chars (${elapsed}s)`);
+
+  return parseJsonResponse(content, queryId);
+}
+
+/**
+ * Sanitize JSON string
+ */
+function sanitizeJson(jsonStr: string): string {
+  let s = jsonStr;
+  s = s.replace(/"amount"\s*:\s*\+/g, '"amount": ');
+  s = s.replace(/:\s*\+(\d)/g, ': $1');
+  s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3.$4');
+  s = s.replace(/,\s*([}\]])/g, '$1');
+  s = s.replace(/"([^"\\]*)\n([^"]*)"/g, '"$1 $2"');
+  s = s.replace(/"([^"\\]*)\t([^"]*)"/g, '"$1 $2"');
+  s = s.replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, ' ');
+  return s;
+}
+
+/**
+ * Parse amount from various formats
+ */
+function parseAmount(value: unknown): number {
+  if (typeof value === 'number') return value;
+  if (typeof value !== 'string') return 0;
+
+  let s = value.replace(/[€$£\s]/g, '').replace('−', '-').replace('–', '-');
+  if (s.includes(',') && s.indexOf(',') > s.lastIndexOf('.')) {
+    s = s.replace(/\./g, '').replace(',', '.');
+  } else {
+    s = s.replace(/,/g, '');
+  }
+  return parseFloat(s) || 0;
+}
+
+/**
+ * Parse JSON response into transactions
+ */
+function parseJsonResponse(response: string, queryId: string): ITransaction[] {
+  // Remove thinking tags if present
+  let cleanResponse = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+
+  // Debug: show what we're working with
+  console.log(`      [${queryId}] Response preview: ${cleanResponse.substring(0, 300)}...`);
+
+  const codeBlockMatch = cleanResponse.match(/```(?:json)?\s*([\s\S]*?)```/);
+  let jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : cleanResponse;
+  jsonStr = sanitizeJson(jsonStr);
+
+  try {
+    const parsed = JSON.parse(jsonStr);
+    if (Array.isArray(parsed)) {
+      const txs = parsed.map(tx => ({
+        date: String(tx.date || ''),
+        counterparty: String(tx.counterparty || tx.description || ''),
+        amount: parseAmount(tx.amount),
+      }));
+      console.log(`      [${queryId}] Parsed ${txs.length} transactions`);
+      return txs;
+    }
+  } catch (e) {
+    // Try to find a JSON array in the text
+    const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
+    if (arrayMatch) {
+      console.log(`      [${queryId}] Array match found: ${arrayMatch[0].length} chars`);
+      try {
+        const parsed = JSON.parse(sanitizeJson(arrayMatch[0]));
+        if (Array.isArray(parsed)) {
+          const txs = parsed.map(tx => ({
+            date: String(tx.date || ''),
+            counterparty: String(tx.counterparty || tx.description || ''),
+            amount: parseAmount(tx.amount),
+          }));
+          console.log(`      [${queryId}] Parsed ${txs.length} transactions (array match)`);
+          return txs;
+        }
+      } catch (innerErr) {
+        console.log(`      [${queryId}] Array parse error: ${(innerErr as Error).message}`);
+      }
+    } else {
+      console.log(`      [${queryId}] No JSON array found in response`);
+    }
+  }
+
+  console.log(`      [${queryId}] PARSE FAILED`);
+  return [];
+}
+
+/**
+ * Extract transactions (single pass)
+ */
+async function extractTransactions(markdown: string, docName: string): Promise<ITransaction[]> {
+  console.log(`    [${docName}] Extracting...`);
+  const txs = await extractTransactionsFromMarkdown(markdown, docName);
+  console.log(`    [${docName}] Extracted ${txs.length} transactions`);
+  return txs;
+}
+
+/**
+ * Compare transactions
+ */
+function compareTransactions(
+  extracted: ITransaction[],
+  expected: ITransaction[]
+): { matches: number; total: number; errors: string[] } {
+  const errors: string[] = [];
+  let matches = 0;
+
+  for (let i = 0; i < expected.length; i++) {
+    const exp = expected[i];
+    const ext = extracted[i];
+
+    if (!ext) {
+      errors.push(`Missing tx ${i}: ${exp.date} ${exp.counterparty}`);
+      continue;
+    }
+
+    const dateMatch = ext.date === exp.date;
+    const amountMatch = Math.abs(ext.amount - exp.amount) < 0.01;
+
+    if (dateMatch && amountMatch) {
+      matches++;
+    } else {
+      errors.push(`Mismatch ${i}: exp ${exp.date}/${exp.amount}, got ${ext.date}/${ext.amount}`);
+    }
+  }
+
+  if (extracted.length > expected.length) {
+    errors.push(`Extra transactions: ${extracted.length - expected.length}`);
+  }
+
+  return { matches, total: expected.length, errors };
+}
+
+/**
+ * Find all test cases
+ */
+function findTestCases(): ITestCase[] {
+  const testDir = path.join(process.cwd(), '.nogit');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: ITestCase[] = [];
+
+  for (const pdf of files.filter((f: string) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${baseName}.json`;
+    if (files.includes(jsonFile)) {
+      testCases.push({
+        name: baseName,
+        pdfPath: path.join(testDir, pdf),
+        jsonPath: path.join(testDir, jsonFile),
+      });
+    }
+  }
+
+  return testCases.sort((a, b) => a.name.localeCompare(b.name));
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} bank statement test cases\n`);
+
+// Ensure temp directory exists
+if (!fs.existsSync(TEMP_MD_DIR)) {
+  fs.mkdirSync(TEMP_MD_DIR, { recursive: true });
+}
+
+// -------- STAGE 1: OCR with Nanonets --------
+
+// Check if all markdown files already exist
+function allMarkdownFilesExist(): boolean {
+  for (const tc of testCases) {
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Track whether we need to run Stage 1
+let stage1Needed = !allMarkdownFilesExist();
+
+tap.test('Stage 1: Setup Nanonets', async () => {
+  console.log('\n========== STAGE 1: Nanonets OCR ==========\n');
+
+  if (!stage1Needed) {
+    console.log('  [SKIP] All markdown files already exist, skipping Nanonets setup');
+    return;
+  }
+
+  const ok = await ensureNanonetsOcr();
+  expect(ok).toBeTrue();
+});
+
+tap.test('Stage 1: Convert all documents to markdown', async () => {
+  if (!stage1Needed) {
+    console.log('  [SKIP] Using existing markdown files from previous run\n');
+    // Load existing markdown paths
+    for (const tc of testCases) {
+      tc.markdownPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+      console.log(`    Loaded: ${tc.markdownPath}`);
+    }
+    return;
+  }
+
+  console.log('\n  Converting all PDFs to markdown with Nanonets-OCR-s...\n');
+
+  for (const tc of testCases) {
+    console.log(`\n    === ${tc.name} ===`);
+
+    // Convert PDF to images
+    const images = convertPdfToImages(tc.pdfPath);
+    console.log(`      Pages: ${images.length}`);
+
+    // Convert to markdown
+    const markdown = await convertDocumentToMarkdown(images, tc.name);
+
+    // Save markdown to temp file
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    fs.writeFileSync(mdPath, markdown);
+    tc.markdownPath = mdPath;
+    console.log(`      Saved: ${mdPath}`);
+  }
+
+  console.log('\n  Stage 1 complete: All documents converted to markdown\n');
+});
+
+tap.test('Stage 1: Stop Nanonets', async () => {
+  if (!stage1Needed) {
+    console.log('  [SKIP] Nanonets was not started');
+    return;
+  }
+
+  stopNanonets();
+  // Verify it's stopped
+  await new Promise(resolve => setTimeout(resolve, 3000));
+  expect(isContainerRunning('nanonets-test')).toBeFalse();
+});
+
+// -------- STAGE 2: Extraction with GPT-OSS 20B --------
+
+tap.test('Stage 2: Setup Ollama + GPT-OSS 20B', async () => {
+  console.log('\n========== STAGE 2: GPT-OSS 20B Extraction ==========\n');
+
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const extractionOk = await ensureExtractionModel();
+  expect(extractionOk).toBeTrue();
+});
+
+let passedCount = 0;
+let failedCount = 0;
+
+for (const tc of testCases) {
+  tap.test(`Stage 2: Extract ${tc.name}`, async () => {
+    const expected: ITransaction[] = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    console.log(`\n    === ${tc.name} ===`);
+    console.log(`    Expected: ${expected.length} transactions`);
+
+    // Load saved markdown
+    const mdPath = path.join(TEMP_MD_DIR, `${tc.name}.md`);
+    if (!fs.existsSync(mdPath)) {
+      throw new Error(`Markdown not found: ${mdPath}. Run Stage 1 first.`);
+    }
+    const markdown = fs.readFileSync(mdPath, 'utf-8');
+    console.log(`    Markdown: ${markdown.length} chars`);
+
+    // Extract transactions (single pass)
+    const extracted = await extractTransactions(markdown, tc.name);
+
+    // Log results
+    console.log(`    Extracted: ${extracted.length} transactions`);
+    for (let i = 0; i < Math.min(extracted.length, 5); i++) {
+      const tx = extracted[i];
+      console.log(`      ${i + 1}. ${tx.date} | ${tx.counterparty.substring(0, 25).padEnd(25)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
+    }
+    if (extracted.length > 5) {
+      console.log(`      ... and ${extracted.length - 5} more`);
+    }
+
+    // Compare
+    const result = compareTransactions(extracted, expected);
+    const pass = result.matches === result.total && extracted.length === expected.length;
+
+    if (pass) {
+      passedCount++;
+      console.log(`    Result: PASS (${result.matches}/${result.total})`);
+    } else {
+      failedCount++;
+      console.log(`    Result: FAIL (${result.matches}/${result.total})`);
+      result.errors.slice(0, 5).forEach(e => console.log(`      - ${e}`));
+    }
+
+    expect(result.matches).toEqual(result.total);
+    expect(extracted.length).toEqual(expected.length);
+  });
+}
+
+tap.test('Summary', async () => {
+  console.log(`\n======================================================`);
+  console.log(`   Bank Statement Summary (Nanonets + GPT-OSS 20B Sequential)`);
+  console.log(`======================================================`);
+  console.log(`  Stage 1:  Nanonets-OCR-s (document -> markdown)`);
+  console.log(`  Stage 2:  GPT-OSS 20B (markdown -> JSON)`);
+  console.log(`  Passed:   ${passedCount}/${testCases.length}`);
+  console.log(`  Failed:   ${failedCount}/${testCases.length}`);
+  console.log(`======================================================\n`);
+
+  // Only cleanup temp files if ALL tests passed
+  if (failedCount === 0 && passedCount === testCases.length) {
+    try {
+      fs.rmSync(TEMP_MD_DIR, { recursive: true, force: true });
+      console.log(`  Cleaned up temp directory: ${TEMP_MD_DIR}\n`);
+    } catch {
+      // Ignore
+    }
+  } else {
+    console.log(`  Keeping temp directory for debugging: ${TEMP_MD_DIR}\n`);
+  }
+});
+
+export default tap.start();

test/test.invoices.nanonets.ts

@@ -0,0 +1,604 @@
+/**
+ * Invoice extraction using Nanonets-OCR-s + Qwen3 (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)
+ *
+ * This approach avoids GPU contention by running services sequentially.
+ */
+import { tap, expect } from '@git.zone/tstest/tapbundle';
+import * as fs from 'fs';
+import * as path from 'path';
+import { execSync } from 'child_process';
+import * as os from 'os';
+import { ensureNanonetsOcr, ensureMiniCpm, isContainerRunning } from './helpers/docker.js';
+
+const NANONETS_URL = 'http://localhost:8000/v1';
+const NANONETS_MODEL = 'nanonets/Nanonets-OCR-s';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const QWEN_MODEL = 'qwen3:8b';
+
+// Temp directory for storing markdown between stages
+const TEMP_MD_DIR = path.join(os.tmpdir(), 'nanonets-invoices-markdown');
+
+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;
+  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 Qwen3
+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)
+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:
+`;
+
+/**
+ * Convert PDF to PNG images
+ */
+function convertPdfToImages(pdfPath: string): string[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
+  const outputPattern = path.join(tempDir, 'page-%d.png');
+
+  try {
+    execSync(
+      `convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
+      { stdio: 'pipe' }
+    );
+
+    const files = fs.readdirSync(tempDir).filter((f) => f.endsWith('.png')).sort();
+    const images: string[] = [];
+
+    for (const file of files) {
+      const imagePath = path.join(tempDir, file);
+      const imageData = fs.readFileSync(imagePath);
+      images.push(imageData.toString('base64'));
+    }
+
+    return images;
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Convert a single page to markdown using Nanonets-OCR-s
+ */
+async function convertPageToMarkdown(image: string, pageNum: number): Promise<string> {
+  const startTime = Date.now();
+
+  const response = await fetch(`${NANONETS_URL}/chat/completions`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': 'Bearer dummy',
+    },
+    body: JSON.stringify({
+      model: NANONETS_MODEL,
+      messages: [{
+        role: 'user',
+        content: [
+          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
+          { type: 'text', text: NANONETS_OCR_PROMPT },
+        ],
+      }],
+      max_tokens: 4096,
+      temperature: 0.0,
+    }),
+  });
+
+  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();
+  const content = (data.choices?.[0]?.message?.content || '').trim();
+  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
+  return content;
+}
+
+/**
+ * Convert all pages of a document to markdown
+ */
+async function convertDocumentToMarkdown(images: string[], docName: string): Promise<string> {
+  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+
+  const markdownPages: string[] = [];
+
+  for (let i = 0; i < images.length; i++) {
+    const markdown = await convertPageToMarkdown(images[i], i + 1);
+    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+  }
+
+  const fullMarkdown = markdownPages.join('\n\n');
+  console.log(`      [${docName}] Complete: ${fullMarkdown.length} chars total`);
+  return fullMarkdown;
+}
+
+/**
+ * Stop Nanonets container
+ */
+function stopNanonets(): void {
+  console.log('    [Docker] Stopping Nanonets container...');
+  try {
+    execSync('docker stop nanonets-test 2>/dev/null || true', { stdio: 'pipe' });
+    execSync('sleep 5', { stdio: 'pipe' });
+    console.log('    [Docker] Nanonets stopped');
+  } catch {
+    console.log('    [Docker] Nanonets was not running');
+  }
+}
+
+/**
+ * Ensure Qwen3 model is available
+ */
+async function ensureQwen3(): Promise<boolean> {
+  try {
+    const response = await fetch(`${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}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`    [Ollama] Pulling ${QWEN_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: QWEN_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+/**
+ * Parse amount from string (handles European format)
+ */
+function parseAmount(s: string | number | undefined): number {
+  if (s === undefined || s === null) return 0;
+  if (typeof s === 'number') return s;
+  const match = s.match(/([\d.,]+)/);
+  if (!match) return 0;
+  const numStr = match[1];
+  const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
+    ? numStr.replace(/\./g, '').replace(',', '.')
+    : numStr.replace(/,/g, '');
+  return parseFloat(normalized) || 0;
+}
+
+/**
+ * Extract invoice number from potentially verbose response
+ */
+function extractInvoiceNumber(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const patterns = [
+    /\b([A-Z]{2,3}\d{10,})\b/i,
+    /\b([A-Z]\d{8,})\b/i,
+    /\b(INV[-\s]?\d{4}[-\s]?\d+)\b/i,
+    /\b(\d{7,})\b/,
+  ];
+  for (const pattern of patterns) {
+    const match = clean.match(pattern);
+    if (match) return match[1];
+  }
+  return clean.replace(/[^A-Z0-9-]/gi, '').trim() || clean;
+}
+
+/**
+ * Extract date (YYYY-MM-DD) from response
+ */
+function extractDate(s: string | undefined): string {
+  if (!s) return '';
+  let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
+  const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
+  if (isoMatch) return isoMatch[1];
+  const dmyMatch = clean.match(/(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})/);
+  if (dmyMatch) {
+    return `${dmyMatch[3]}-${dmyMatch[2].padStart(2, '0')}-${dmyMatch[1].padStart(2, '0')}`;
+  }
+  return clean.replace(/[^\d-]/g, '').trim();
+}
+
+/**
+ * Extract currency
+ */
+function extractCurrency(s: string | undefined): string {
+  if (!s) return 'EUR';
+  const upper = s.toUpperCase();
+  if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
+  if (upper.includes('USD') || upper.includes('$')) return 'USD';
+  if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
+  return 'EUR';
+}
+
+/**
+ * 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 Qwen3
+ */
+async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Promise<IInvoice | null> {
+  console.log(`      [${queryId}] Sending to ${QWEN_MODEL}...`);
+  const startTime = Date.now();
+
+  const response = await fetch(`${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,
+      },
+    }),
+  });
+
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+
+  if (!response.ok) {
+    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)`);
+
+  return parseJsonToInvoice(content);
+}
+
+/**
+ * Compare two invoices for consensus
+ */
+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} | ${fallback.total_amount}`);
+    return fallback;
+  }
+
+  return {
+    invoice_number: '',
+    invoice_date: '',
+    vendor_name: '',
+    currency: 'EUR',
+    net_amount: 0,
+    vat_amount: 0,
+    total_amount: 0,
+  };
+}
+
+/**
+ * Normalize date to YYYY-MM-DD
+ */
+function normalizeDate(dateStr: string | null): string {
+  if (!dateStr) return '';
+  if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
+
+  const monthMap: Record<string, string> = {
+    JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
+    JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
+  };
+
+  let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
+  if (match) {
+    return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
+  }
+
+  match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
+  if (match) {
+    return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
+  }
+
+  return dateStr;
+}
+
+/**
+ * Compare extracted invoice against expected
+ */
+function compareInvoice(
+  extracted: IInvoice,
+  expected: IInvoice
+): { match: boolean; errors: string[] } {
+  const errors: string[] = [];
+
+  const extNum = extracted.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
+  const expNum = expected.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
+  if (extNum !== expNum) {
+    errors.push(`invoice_number: exp "${expected.invoice_number}", got "${extracted.invoice_number}"`);
+  }
+
+  if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
+    errors.push(`invoice_date: exp "${expected.invoice_date}", got "${extracted.invoice_date}"`);
+  }
+
+  if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
+    errors.push(`total_amount: exp ${expected.total_amount}, got ${extracted.total_amount}`);
+  }
+
+  if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
+    errors.push(`currency: exp "${expected.currency}", got "${extracted.currency}"`);
+  }
+
+  return { match: errors.length === 0, errors };
+}
+
+/**
+ * Find all test cases
+ */
+function findTestCases(): ITestCase[] {
+  const testDir = path.join(process.cwd(), '.nogit/invoices');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: ITestCase[] = [];
+
+  for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${baseName}.json`;
+    if (files.includes(jsonFile)) {
+      testCases.push({
+        name: baseName,
+        pdfPath: path.join(testDir, pdf),
+        jsonPath: path.join(testDir, jsonFile),
+      });
+    }
+  }
+
+  return testCases.sort((a, b) => a.name.localeCompare(b.name));
+}
+
+// ============ TESTS ============
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} invoice test cases\n`);
+
+// Ensure 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 all invoices to markdown', async () => {
+  console.log('\n  Converting all 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}`);
+  }
+
+  console.log('\n  Stage 1 complete: All 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 Qwen3 --------
+
+tap.test('Stage 2: Setup Ollama + Qwen3', async () => {
+  console.log('\n========== STAGE 2: Qwen3 Extraction ==========\n');
+
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  const qwenOk = await ensureQwen3();
+  expect(qwenOk).toBeTrue();
+});
+
+let passedCount = 0;
+let failedCount = 0;
+const processingTimes: number[] = [];
+
+for (const tc of testCases) {
+  tap.test(`Stage 2: Extract ${tc.name}`, async () => {
+    const expected: IInvoice = JSON.parse(fs.readFileSync(tc.jsonPath, 'utf-8'));
+    console.log(`\n    === ${tc.name} ===`);
+    console.log(`    Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
+
+    const startTime = Date.now();
+
+    const mdPath = path.join(TEMP_MD_DIR, `${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 extractWithConsensus(markdown, tc.name);
+
+    const elapsedMs = Date.now() - startTime;
+    processingTimes.push(elapsedMs);
+
+    console.log(`    Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
+
+    const result = compareInvoice(extracted, expected);
+
+    if (result.match) {
+      passedCount++;
+      console.log(`    Result: MATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+    } else {
+      failedCount++;
+      console.log(`    Result: MISMATCH (${(elapsedMs / 1000).toFixed(1)}s)`);
+      result.errors.forEach(e => console.log(`      - ${e}`));
+    }
+
+    expect(result.match).toBeTrue();
+  });
+}
+
+tap.test('Summary', async () => {
+  const totalInvoices = testCases.length;
+  const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
+  const totalTimeMs = processingTimes.reduce((a, b) => a + b, 0);
+  const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
+
+  console.log(`\n========================================`);
+  console.log(`   Invoice Summary (Nanonets + Qwen3)`);
+  console.log(`========================================`);
+  console.log(`  Stage 1:   Nanonets-OCR-s (doc -> md)`);
+  console.log(`  Stage 2:   Qwen3 8B (md -> JSON)`);
+  console.log(`  Passed:    ${passedCount}/${totalInvoices}`);
+  console.log(`  Failed:    ${failedCount}/${totalInvoices}`);
+  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
+  console.log(`----------------------------------------`);
+  console.log(`  Total time:   ${(totalTimeMs / 1000).toFixed(1)}s`);
+  console.log(`  Avg per inv:  ${avgTimeSec.toFixed(1)}s`);
+  console.log(`========================================\n`);
+
+  // Cleanup temp files
+  try {
+    fs.rmSync(TEMP_MD_DIR, { recursive: true, force: true });
+    console.log(`  Cleaned up temp directory: ${TEMP_MD_DIR}\n`);
+  } catch {
+    // Ignore
+  }
+});
+
+export default tap.start();