update

readme.md

@@ -1,19 +1,27 @@
 # @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 or CPU fallback—**no cloud API keys required**.
+
+> 🔥 **Four 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 |
+| Model | Parameters | Best For | API | Port |
-|-------|-----------|----------|-----|
+|-------|-----------|----------|-----|------|
-| **MiniCPM-V 4.5** | 8B | General vision understanding, image analysis, multi-image | Ollama-compatible |
+| **MiniCPM-V 4.5** | 8B | General vision understanding, multi-image analysis | Ollama-compatible | 11434 |
-| **PaddleOCR-VL** | 0.9B | Document parsing, table extraction, OCR | OpenAI-compatible |
+| **PaddleOCR-VL** | 0.9B | Document parsing, table extraction, structured OCR | OpenAI-compatible | 8000 |
+| **Nanonets-OCR-s** | ~4B | Document OCR with semantic markdown output | OpenAI-compatible | 8000 |
+| **Qwen3-VL-30B** | 30B (A3B) | Advanced visual agents, code generation from images | Ollama-compatible | 11434 |
 
-## 📦 Available Images
+---
+
+## 📦 Quick Reference: All Available Images
 
 ```
 code.foss.global/host.today/ht-docker-ai:<tag>
@@ -25,12 +33,14 @@ code.foss.global/host.today/ht-docker-ai:<tag>
 | `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 |
+| `nanonets-ocr` | Nanonets-OCR-s | NVIDIA GPU (8-10GB VRAM) | 8000 |
+| `qwen3vl` | Qwen3-VL-30B-A3B | NVIDIA GPU (~20GB VRAM) | 11434 |
 
 ---
 
 ## 🖼️ 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
 
@@ -95,7 +105,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ## 📄 PaddleOCR-VL
 
-A specialized 0.9B Vision-Language Model optimized for document parsing. Native support for tables, formulas, charts, and text extraction in 109 languages.
+A specialized **0.9B Vision-Language Model** optimized for document parsing. Native support for tables, formulas, charts, and text extraction in **109 languages**.
 
 ### Quick Start
 
@@ -185,8 +195,121 @@ PaddleOCR-VL accepts images in multiple formats:
 
 ---
 
+## 🔍 Nanonets-OCR-s
+
+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:** ~8-10GB VRAM, runs great on consumer GPUs
+- 🔌 **OpenAI-compatible:** Drop-in replacement for existing pipelines
+
+### Quick Start
+
+```bash
+docker run -d \
+  --name nanonets \
+  --gpus all \
+  -p 8000:8000 \
+  -v hf-cache:/root/.cache/huggingface \
+  code.foss.global/host.today/ht-docker-ai:nanonets-ocr
+```
+
+### API Usage
+
+```bash
+curl http://localhost:8000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "nanonets/Nanonets-OCR-s",
+    "messages": [{
+      "role": "user",
+      "content": [
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,<base64>"}},
+        {"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."}
+      ]
+    }],
+    "temperature": 0.0,
+    "max_tokens": 4096
+  }'
+```
+
+### Output Format
+
+Nanonets-OCR-s returns markdown with semantic tags:
+
+| 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
+
+| Metric | Value |
+|--------|-------|
+| Speed | 3–8 seconds per page |
+| VRAM | ~8-10GB |
+
+---
+
+## 🧠 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:
@@ -206,7 +329,7 @@ services:
               capabilities: [gpu]
     restart: unless-stopped
 
-  # Document parsing / OCR
+  # Document parsing / OCR (table specialist)
   paddleocr:
     image: code.foss.global/host.today/ht-docker-ai:paddleocr-vl
     ports:
@@ -222,6 +345,22 @@ services:
               capabilities: [gpu]
     restart: unless-stopped
 
+  # Document OCR with semantic output
+  nanonets:
+    image: code.foss.global/host.today/ht-docker-ai:nanonets-ocr
+    ports:
+      - "8001:8000"
+    volumes:
+      - hf-cache:/root/.cache/huggingface
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    restart: unless-stopped
+
 volumes:
   ollama-data:
   hf-cache:
@@ -231,7 +370,7 @@ volumes:
 
 ## ⚙️ Environment Variables
 
-### MiniCPM-V 4.5
+### MiniCPM-V 4.5 & Qwen3-VL (Ollama-based)
 
 | Variable | Default | Description |
 |----------|---------|-------------|
@@ -239,13 +378,47 @@ volumes:
 | `OLLAMA_HOST` | `0.0.0.0` | API bind address |
 | `OLLAMA_ORIGINS` | `*` | Allowed CORS origins |
 
-### PaddleOCR-VL
+### PaddleOCR-VL & Nanonets-OCR (vLLM-based)
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `MODEL_NAME` | `PaddlePaddle/PaddleOCR-VL` | HuggingFace model ID |
+| `MODEL_NAME` | Model-specific | HuggingFace model ID |
-| `SERVER_HOST` | `0.0.0.0` | API bind address |
+| `HOST` | `0.0.0.0` | API bind address |
-| `SERVER_PORT` | `8000` | API port |
+| `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:** 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 Multi-Model Works
+
+- **Different architectures:** Independent models cross-validate each other
+- **Specialized strengths:** PaddleOCR-VL excels at tables; 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 |
+| Table extraction from documents | PaddleOCR-VL |
+| 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 +438,16 @@ 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 +467,13 @@ sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 ```
 
+### GPU Memory Contention (Multi-Model)
+
+When running multiple VLMs on a single GPU:
+- vLLM and Ollama both need significant GPU memory
+- **Single GPU:** Run services sequentially (stop one before starting another)
+- **Multi-GPU:** Assign each service to a different GPU via `CUDA_VISIBLE_DEVICES`
+
 ---
 
 ## License and Legal Information

test/test.bankstatements.nanonets.ts

@@ -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 +57,48 @@ 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));
+}
+
+/**
+ * Batch images to fit within context window
+ */
+function batchImages(images: IImageData[]): IImageData[][] {
+  const batches: IImageData[][] = [];
+  let currentBatch: IImageData[] = [];
+  let currentTokens = 0;
+
+  for (const img of images) {
+    const imgTokens = estimateVisualTokens(img.width, img.height);
+
+    if (currentTokens + imgTokens > MAX_VISUAL_TOKENS && currentBatch.length > 0) {
+      batches.push(currentBatch);
+      currentBatch = [img];
+      currentTokens = imgTokens;
+    } else {
+      currentBatch.push(img);
+      currentTokens += imgTokens;
+    }
+  }
+  if (currentBatch.length > 0) batches.push(currentBatch);
+
+  return batches;
+}
+
+/**
+ * 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 +106,24 @@ function convertPdfToImages(pdfPath: string): string[] {
       { stdio: 'pipe' }
     );
 
-    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.png')).sort();
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
-    const images: string[] = [];
+    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 +133,28 @@ 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(`${NANONETS_URL}/chat/completions`, {
     method: 'POST',
@@ -94,12 +166,9 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
       model: NANONETS_MODEL,
       messages: [{
         role: 'user',
-        content: [
+        content,
-          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
-          { type: 'text', text: NANONETS_OCR_PROMPT },
-        ],
       }],
-      max_tokens: 4096,
+      max_tokens: 4096 * batch.length,  // Scale output tokens with batch size
       temperature: 0.0,
     }),
   });
@@ -112,25 +181,35 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
   }
 
   const data = await response.json();
-  const content = (data.choices?.[0]?.message?.content || '').trim();
+  let responseContent = (data.choices?.[0]?.message?.content || '').trim();
-  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
+
-  return content;
+  // 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> {
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
-  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+  const batches = batchImages(images);
+  console.log(`      [${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++) {
+  for (let i = 0; i < batches.length; i++) {
-    const markdown = await convertPageToMarkdown(images[i], i + 1);
+    const batch = batches[i];
-    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+    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 +240,6 @@ 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 (${warmupData.message?.content?.length || 0} chars)`);
-        }
-
         return true;
       }
     }
@@ -201,22 +261,24 @@ async function ensureExtractionModel(): Promise<boolean> {
  * Extract transactions from markdown using GPT-OSS 20B (streaming)
  */
 async function extractTransactionsFromMarkdown(markdown: string, queryId: string): Promise<ITransaction[]> {
-  console.log(`      [${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: [{
+      messages: [
-        role: 'user',
+        { role: 'user', content: 'Hi there, how are you?' },
-        content: fullPrompt,
+        { 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 +290,45 @@ 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 +337,12 @@ 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);
 }

test/test.invoices.nanonets.ts

@@ -1,8 +1,8 @@
 /**
- * Invoice extraction using Nanonets-OCR-s + Qwen3 (sequential two-stage pipeline)
+ * Invoice 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: Qwen3 extracts structured JSON from saved markdown (after Nanonets stops)
+ * Stage 2: GPT-OSS 20B extracts structured JSON from saved markdown (after Nanonets stops)
  *
  * This approach avoids GPU contention by running services sequentially.
  */
@@ -17,7 +17,7 @@ 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';
+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 +80,48 @@ 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));
+}
+
+/**
+ * Batch images to fit within context window
+ */
+function batchImages(images: IImageData[]): IImageData[][] {
+  const batches: IImageData[][] = [];
+  let currentBatch: IImageData[] = [];
+  let currentTokens = 0;
+
+  for (const img of images) {
+    const imgTokens = estimateVisualTokens(img.width, img.height);
+
+    if (currentTokens + imgTokens > MAX_VISUAL_TOKENS && currentBatch.length > 0) {
+      batches.push(currentBatch);
+      currentBatch = [img];
+      currentTokens = imgTokens;
+    } else {
+      currentBatch.push(img);
+      currentTokens += imgTokens;
+    }
+  }
+  if (currentBatch.length > 0) batches.push(currentBatch);
+
+  return batches;
+}
+
+/**
+ * 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 +129,24 @@ function convertPdfToImages(pdfPath: string): string[] {
       { stdio: 'pipe' }
     );
 
-    const files = fs.readdirSync(tempDir).filter((f) => f.endsWith('.png')).sort();
+    const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.jpg')).sort();
-    const images: string[] = [];
+    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 +156,28 @@ 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(`${NANONETS_URL}/chat/completions`, {
     method: 'POST',
@@ -117,12 +189,9 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
       model: NANONETS_MODEL,
       messages: [{
         role: 'user',
-        content: [
+        content,
-          { type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
-          { type: 'text', text: NANONETS_OCR_PROMPT },
-        ],
       }],
-      max_tokens: 4096,
+      max_tokens: 4096 * batch.length,  // Scale output tokens with batch size
       temperature: 0.0,
     }),
   });
@@ -135,25 +204,35 @@ async function convertPageToMarkdown(image: string, pageNum: number): Promise<st
   }
 
   const data = await response.json();
-  const content = (data.choices?.[0]?.message?.content || '').trim();
+  let responseContent = (data.choices?.[0]?.message?.content || '').trim();
-  console.log(`        Page ${pageNum}: ${content.length} chars (${elapsed}s)`);
+
-  return content;
+  // 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> {
+async function convertDocumentToMarkdown(images: IImageData[], docName: string): Promise<string> {
-  console.log(`      [${docName}] Converting ${images.length} page(s)...`);
+  const batches = batchImages(images);
+  console.log(`      [${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++) {
+  for (let i = 0; i < batches.length; i++) {
-    const markdown = await convertPageToMarkdown(images[i], i + 1);
+    const batch = batches[i];
-    markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
+    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 +252,16 @@ 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)) {
+      if (models.some((m: { name: string }) => m.name === EXTRACTION_MODEL)) {
-        console.log(`    [Ollama] Model available: ${QWEN_MODEL}`);
+        console.log(`    [Ollama] Model available: ${EXTRACTION_MODEL}`);
         return true;
       }
     }
@@ -190,11 +269,11 @@ 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 +382,102 @@ 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,
+      model: EXTRACTION_MODEL,
-      messages: [{
+      messages: [
-        role: 'user',
+        { role: 'user', content: 'Hi there, how are you?' },
-        content: JSON_EXTRACTION_PROMPT + markdown,
+        { role: 'assistant', content: 'Good, how can I help you today?' },
-      }],
+        { role: 'user', content: fullPrompt },
-      stream: false,
+      ],
-      options: {
+      stream: true,
-        num_predict: 2000,
-        temperature: 0.1,
-      },
     }),
+    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();
+  // Stream the response
-  const content = (data.message?.content || '').trim();
+  let content = '';
-  console.log(`      [${queryId}] Response: ${content.length} chars (${elapsed}s)`);
+  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 {
+async function extractInvoice(markdown: string, docName: string): Promise<IInvoice> {
-  const numMatch = a.invoice_number.toLowerCase() === b.invoice_number.toLowerCase();
+  console.log(`    [${docName}] Extracting...`);
-  const dateMatch = a.invoice_date === b.invoice_date;
+  const invoice = await extractInvoiceFromMarkdown(markdown, docName);
-  const totalMatch = Math.abs(a.total_amount - b.total_amount) < 0.02;
+  if (!invoice) {
-  return numMatch && dateMatch && totalMatch;
-}
-
-/**
- * Extract with consensus
- */
-async function extractWithConsensus(markdown: string, docName: string): Promise<IInvoice> {
-  const MAX_ATTEMPTS = 3;
-
-  for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt++) {
-    console.log(`    [${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} | ${fallback.total_amount}`);
-    return fallback;
-  }
-
     return {
       invoice_number: '',
       invoice_date: '',
@@ -394,6 +487,9 @@ async function extractWithConsensus(markdown: string, docName: string): Promise<
       vat_amount: 0,
       total_amount: 0,
     };
+  }
+  console.log(`    [${docName}] Extracted: ${invoice.invoice_number}`);
+  return invoice;
 }
 
 /**
@@ -520,16 +616,16 @@ 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 () => {
+tap.test('Stage 2: Setup Ollama + GPT-OSS 20B', async () => {
-  console.log('\n========== STAGE 2: Qwen3 Extraction ==========\n');
+  console.log('\n========== STAGE 2: GPT-OSS 20B Extraction ==========\n');
 
   const ollamaOk = await ensureMiniCpm();
   expect(ollamaOk).toBeTrue();
 
-  const qwenOk = await ensureQwen3();
+  const extractionOk = await ensureExtractionModel();
-  expect(qwenOk).toBeTrue();
+  expect(extractionOk).toBeTrue();
 });
 
 let passedCount = 0;
@@ -551,7 +647,7 @@ 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 +676,10 @@ 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)}%`);