v1.14.2

changelog.md

@@ -1,5 +1,24 @@
 # Changelog
 
+## 2026-01-19 - 1.14.2 - fix(readme)
+update README to document Nanonets-OCR2-3B (replaces Nanonets-OCR-s), adjust VRAM and context defaults, expand feature docs, and update examples/test command
+
+- Renamed Nanonets-OCR-s -> Nanonets-OCR2-3B throughout README and examples
+- Updated Nanonets VRAM guidance from ~10GB to ~12-16GB and documented 30K context
+- Changed documented MAX_MODEL_LEN default from 8192 to 30000
+- Updated example model identifiers (model strings and curl/example snippets) to nanonets/Nanonets-OCR2-3B
+- Added MiniCPM and Qwen feature bullets (multilingual, multi-image, flowchart support, expanded context notes)
+- Replaced README test command from ./test-images.sh to pnpm test
+
+## 2026-01-19 - 1.14.1 - fix(extraction)
+improve JSON extraction prompts and model options for invoice and bank statement tests
+
+- Refactor JSON extraction prompts to be sent after the document text and add explicit 'WHERE TO FIND DATA' and 'RULES' sections for clearer extraction guidance
+- Change chat message flow to: send document, assistant acknowledgement, then the JSON extraction prompt (avoids concatenating large prompts into one message)
+- Add model options (num_ctx: 32768, temperature: 0) to give larger context windows and deterministic JSON output
+- Simplify logging to avoid printing full prompt contents; log document and prompt lengths instead
+- Increase timeouts for large documents to 600000ms (10 minutes) where applicable
+
 ## 2026-01-19 - 1.14.0 - feat(docker-images)
 add vLLM-based Nanonets-OCR2-3B image, Qwen3-VL Ollama image and refactor build/docs/tests to use new runtime/layout
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@host.today/ht-docker-ai",
-  "version": "1.14.0",
+  "version": "1.14.2",
   "type": "module",
   "private": false,
   "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",

readme.md

@@ -2,7 +2,7 @@
 
 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.
+> 🔥 **Three VLMs, one registry.** From high-performance document OCR to GPT-4o-level vision understanding—pick the right tool for your task.
 
 ## Issue Reporting and Security
 
@@ -15,7 +15,7 @@ For reporting bugs, issues, or security vulnerabilities, please visit [community
 | 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 |
+| **Nanonets-OCR2-3B** | ~3B | Document OCR with semantic markdown, LaTeX, flowcharts | OpenAI-compatible | 8000 | ~12-16GB |
 | **Qwen3-VL-30B** | 30B (A3B) | Advanced visual agents, code generation from images | Ollama-compatible | 11434 | ~20GB |
 
 ---
@@ -29,7 +29,7 @@ code.foss.global/host.today/ht-docker-ai:<tag>
 | Tag | Model | Runtime | Port | VRAM |
 |-----|-------|---------|------|------|
 | `minicpm45v` / `latest` | MiniCPM-V 4.5 | Ollama | 11434 | ~9GB |
-| `nanonets-ocr` | Nanonets-OCR-s | vLLM | 8000 | ~10GB |
+| `nanonets-ocr` | Nanonets-OCR2-3B | vLLM | 8000 | ~12-16GB |
 | `qwen3vl` | Qwen3-VL-30B-A3B | Ollama | 11434 | ~20GB |
 
 ---
@@ -38,6 +38,13 @@ code.foss.global/host.today/ht-docker-ai:<tag>
 
 A GPT-4o level multimodal LLM from OpenBMB—handles image understanding, OCR, multi-image analysis, and visual reasoning across **30+ languages**.
 
+### ✨ Key Features
+
+- 🌍 **Multilingual:** 30+ languages supported
+- 🖼️ **Multi-image:** Analyze multiple images in one request
+- 📊 **Versatile:** Charts, documents, photos, diagrams
+- ⚡ **Efficient:** Runs on consumer GPUs (9GB VRAM)
+
 ### Quick Start
 
 ```bash
@@ -83,21 +90,22 @@ curl http://localhost:11434/api/chat -d '{
 
 | Mode | VRAM Required |
 |------|---------------|
-| int4 quantized | 9GB |
-| Full precision (bf16) | 18GB |
+| int4 quantized | ~9GB |
+| Full precision (bf16) | ~18GB |
 
 ---
 
-## 🔍 Nanonets-OCR-s
+## 🔍 Nanonets-OCR2-3B
 
-A **Qwen2.5-VL-3B** model fine-tuned specifically for document OCR. Outputs structured markdown with semantic HTML tags—perfect for preserving document structure.
+The **latest Nanonets document OCR model** (October 2025 release)—based on Qwen2.5-VL-3B, fine-tuned specifically for document extraction with significant improvements over the original OCR-s.
 
-### Key Features
+### ✨ Key Features
 
-- 📝 **Semantic output:** Tables → HTML, equations → LaTeX, watermarks/page numbers → tagged
+- 📝 **Semantic output:** Tables → HTML, equations → LaTeX, flowcharts → structured markup
 - 🌍 **Multilingual:** Inherits Qwen's broad language support
-- ⚡ **Efficient:** ~10GB VRAM, runs great on consumer GPUs
+- 📄 **30K context:** Handle large, multi-page documents
 - 🔌 **OpenAI-compatible:** Drop-in replacement for existing pipelines
+- 🎯 **Improved accuracy:** Better semantic tagging and LaTeX equation extraction vs. OCR-s
 
 ### Quick Start
 
@@ -116,7 +124,7 @@ docker run -d \
 curl http://localhost:8000/v1/chat/completions \
   -H "Content-Type: application/json" \
   -d '{
-    "model": "nanonets/Nanonets-OCR-s",
+    "model": "nanonets/Nanonets-OCR2-3B",
     "messages": [{
       "role": "user",
       "content": [
@@ -131,7 +139,7 @@ curl http://localhost:8000/v1/chat/completions \
 
 ### Output Format
 
-Nanonets-OCR-s returns markdown with semantic tags:
+Nanonets-OCR2-3B returns markdown with semantic tags:
 
 | Element | Output Format |
 |---------|---------------|
@@ -140,13 +148,14 @@ Nanonets-OCR-s returns markdown with semantic tags:
 | Images | `<img>description</img>` |
 | Watermarks | `<watermark>OFFICIAL COPY</watermark>` |
 | Page numbers | `<page_number>14</page_number>` |
+| Flowcharts | Structured markup |
 
-### Performance
+### Hardware Requirements
 
-| Metric | Value |
-|--------|-------|
-| Speed | 3–8 seconds per page |
-| VRAM | ~10GB |
+| Config | VRAM |
+|--------|------|
+| 30K context (default) | ~12-16GB |
+| Speed | ~3-8 seconds per page |
 
 ---
 
@@ -154,7 +163,7 @@ Nanonets-OCR-s returns markdown with semantic tags:
 
 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
+### ✨ Key Features
 
 - 🚀 **256K context** (expandable to 1M tokens!)
 - 🤖 **Visual agent capabilities** — can plan and execute multi-step tasks
@@ -204,7 +213,6 @@ curl http://localhost:11434/api/chat -d '{
 Run multiple VLMs together for maximum flexibility:
 
 ```yaml
-version: '3.8'
 services:
   # General vision tasks
   minicpm:
@@ -259,10 +267,10 @@ volumes:
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `MODEL_NAME` | `nanonets/Nanonets-OCR-s` | HuggingFace model ID |
+| `MODEL_NAME` | `nanonets/Nanonets-OCR2-3B` | HuggingFace model ID |
 | `HOST` | `0.0.0.0` | API bind address |
 | `PORT` | `8000` | API port |
-| `MAX_MODEL_LEN` | `8192` | Maximum sequence length |
+| `MAX_MODEL_LEN` | `30000` | Maximum sequence length |
 | `GPU_MEMORY_UTILIZATION` | `0.9` | GPU memory usage (0-1) |
 
 ---
@@ -283,7 +291,7 @@ 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
+- **Specialized strengths:** Nanonets-OCR2-3B excels at document structure; MiniCPM-V handles general vision
 - **Native processing:** All VLMs see original images—no intermediate structure loss
 
 ### Model Selection Guide
@@ -291,10 +299,11 @@ This dual-VLM approach catches extraction errors that single models miss.
 | Task | Recommended Model |
 |------|-------------------|
 | General image understanding | MiniCPM-V 4.5 |
-| Document OCR with structure preservation | Nanonets-OCR-s |
+| Document OCR with structure preservation | Nanonets-OCR2-3B |
 | Complex visual reasoning / code generation | Qwen3-VL-30B |
 | Multi-image analysis | MiniCPM-V 4.5 |
 | Visual agent tasks | Qwen3-VL-30B |
+| Large documents (30K+ tokens) | Nanonets-OCR2-3B |
 
 ---
 
@@ -309,7 +318,7 @@ cd ht-docker-ai
 ./build-images.sh
 
 # Run tests
-./test-images.sh
+pnpm test
 ```
 
 ---

test/test.bankstatements.nanonets.ts

@@ -51,11 +51,21 @@ 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 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.
+// JSON extraction prompt for GPT-OSS 20B (sent AFTER the statement text is provided)
+const JSON_EXTRACTION_PROMPT = `Extract ALL transactions from the bank statement. Return ONLY valid JSON array.
 
-STATEMENT:
-`;
+WHERE TO FIND DATA:
+- Transactions are typically in TABLES with columns: Date, Description/Counterparty, Debit, Credit, Balance
+- Look for rows with actual money movements, NOT header rows or summary totals
+
+RULES:
+1. date: Convert to YYYY-MM-DD format
+2. counterparty: The name/description of who the money went to/from
+3. amount: NEGATIVE for debits/withdrawals, POSITIVE for credits/deposits
+4. Only include actual transactions, NOT opening/closing balances
+
+JSON array only:
+[{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-25.99}]`;
 
 // Constants for smart batching
 const MAX_VISUAL_TOKENS = 28000;  // ~32K context minus prompt/output headroom
@@ -246,12 +256,8 @@ async function ensureExtractionModel(): Promise<boolean> {
  */
 async function extractTransactionsFromMarkdown(markdown: string, queryId: string): Promise<ITransaction[]> {
   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 (${fullPrompt.length} chars) =====\n`);
+  console.log(`      [${queryId}] Statement: ${markdown.length} chars, Prompt: ${JSON_EXTRACTION_PROMPT.length} chars`);
 
   const response = await fetch(`${OLLAMA_URL}/api/chat`, {
     method: 'POST',
@@ -261,9 +267,15 @@ async function extractTransactionsFromMarkdown(markdown: string, queryId: string
       messages: [
         { role: 'user', content: 'Hi there, how are you?' },
         { role: 'assistant', content: 'Good, how can I help you today?' },
-        { role: 'user', content: fullPrompt },
+        { role: 'user', content: `Here is a bank statement document:\n\n${markdown}` },
+        { role: 'assistant', content: 'I have read the bank statement document you provided. I can see all the transaction data. What would you like me to do with it?' },
+        { role: 'user', content: JSON_EXTRACTION_PROMPT },
       ],
       stream: true,
+      options: {
+        num_ctx: 32768,  // Larger context for long statements + thinking
+        temperature: 0,  // Deterministic for JSON extraction
+      },
     }),
     signal: AbortSignal.timeout(600000),  // 10 minute timeout
   });

test/test.invoices.extraction.ts

@@ -197,6 +197,10 @@ async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Pr
         { role: 'user', content: JSON_EXTRACTION_PROMPT },
       ],
       stream: true,
+      options: {
+        num_ctx: 32768,  // Larger context for long invoices + thinking
+        temperature: 0,  // Deterministic for JSON extraction
+      },
     }),
     signal: AbortSignal.timeout(120000), // 2 min timeout
   });

test/test.invoices.nanonets.ts

@@ -54,31 +54,24 @@ 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 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.
+// 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.
 
-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
+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: Amount before tax
-6. vat_amount: Tax/VAT amount
-7. total_amount: Final total (gross amount)
+5. net_amount: Total before tax
+6. vat_amount: Tax amount
+7. total_amount: Final total with tax
 
-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:
-`;
+JSON only:
+{"invoice_number":"X","invoice_date":"YYYY-MM-DD","vendor_name":"X","currency":"EUR","net_amount":0,"vat_amount":0,"total_amount":0}`;
 
 // Constants for smart batching
 const MAX_VISUAL_TOKENS = 28000;  // ~32K context minus prompt/output headroom
@@ -370,12 +363,8 @@ function parseJsonToInvoice(response: string): IInvoice | null {
  */
 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 (${fullPrompt.length} chars) =====\n`);
+  console.log(`      [${queryId}] Invoice: ${markdown.length} chars, Prompt: ${JSON_EXTRACTION_PROMPT.length} chars`);
 
   const response = await fetch(`${OLLAMA_URL}/api/chat`, {
     method: 'POST',
@@ -385,9 +374,15 @@ async function extractInvoiceFromMarkdown(markdown: string, queryId: string): Pr
       messages: [
         { role: 'user', content: 'Hi there, how are you?' },
         { role: 'assistant', content: 'Good, how can I help you today?' },
-        { role: 'user', content: fullPrompt },
+        { role: 'user', content: `Here is an invoice document:\n\n${markdown}` },
+        { role: 'assistant', content: 'I have read the invoice document you provided. I can see all the text content. What would you like me to do with it?' },
+        { role: 'user', content: JSON_EXTRACTION_PROMPT },
       ],
       stream: true,
+      options: {
+        num_ctx: 32768,  // Larger context for long invoices + thinking
+        temperature: 0,  // Deterministic for JSON extraction
+      },
     }),
     signal: AbortSignal.timeout(600000), // 10 minute timeout for large documents
   });