v1.10.1

Dockerfile_qwen3vl

@@ -0,0 +1,26 @@
+# Qwen3-VL-30B-A3B Vision Language Model
+# Q4_K_M quantization (~20GB model)
+#
+# Most powerful Qwen vision model:
+# - 256K context (expandable to 1M)
+# - Visual agent capabilities
+# - Code generation from images
+#
+# Build: docker build -f Dockerfile_qwen3vl -t qwen3vl .
+# Run:   docker run --gpus all -p 11434:11434 -v ht-ollama-models:/root/.ollama qwen3vl
+
+FROM ollama/ollama:latest
+
+# Pre-pull the model during build (optional - can also pull at runtime)
+# This makes the image larger but faster to start
+# RUN ollama serve & sleep 5 && ollama pull qwen3-vl:30b-a3b && pkill ollama
+
+# Expose Ollama API port
+EXPOSE 11434
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
+  CMD curl -f http://localhost:11434/api/tags || exit 1
+
+# Start Ollama server
+CMD ["serve"]

changelog.md

@@ -1,5 +1,23 @@
 # Changelog
 
+## 2026-01-18 - 1.10.1 - fix(tests)
+improve Qwen3-VL invoice extraction test by switching to non-stream API, adding model availability/pull checks, simplifying response parsing, and tightening model options
+
+- Replaced streaming reader logic with direct JSON parsing of the /api/chat response
+- Added ensureQwen3Vl() to check and pull the Qwen3-VL:8b model from Ollama
+- Switched to ensureMiniCpm() to verify Ollama service is running before model checks
+- Use /no_think prompt for direct JSON output and set temperature to 0.0 and num_predict to 512
+- Removed retry loop and streaming parsing; improved error messages to include response body
+- Updated logging and test setup messages for clarity
+
+## 2026-01-18 - 1.10.0 - feat(vision)
+add Qwen3-VL vision model support with Dockerfile and tests; improve invoice OCR conversion and prompts; simplify extraction flow by removing consensus voting
+
+- Add Dockerfile_qwen3vl to provide an Ollama-based image for Qwen3-VL and expose the Ollama API on port 11434
+- Introduce test/test.invoices.qwen3vl.ts and ensureQwen3Vl() helper to pull and test qwen3-vl:8b
+- Improve PDF->PNG conversion and prompt in ministral3 tests (higher DPI, max quality, sharpen) and increase num_predict from 512 to 1024
+- Simplify extraction pipeline: remove consensus voting, log single-pass results, and simplify OCR HTML sanitization/truncation logic
+
 ## 2026-01-18 - 1.9.0 - feat(tests)
 add Ministral 3 vision tests and improve invoice extraction pipeline to use Ollama chat schema, sanitization, and multi-page support
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@host.today/ht-docker-ai",
-  "version": "1.9.0",
+  "version": "1.10.1",
   "type": "module",
   "private": false,
   "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",

test/helpers/docker.ts

@@ -311,9 +311,8 @@ export async function ensureOllamaModel(modelName: string): Promise<boolean> {
     if (response.ok) {
       const data = await response.json();
       const models = data.models || [];
-      const exists = models.some((m: { name: string }) =>
-        m.name === modelName || m.name.startsWith(modelName.split(':')[0])
-      );
+      // Exact match required - don't match on prefix
+      const exists = models.some((m: { name: string }) => m.name === modelName);
 
       if (exists) {
         console.log(`[Ollama] Model already available: ${modelName}`);
@@ -371,3 +370,16 @@ export async function ensureMinistral3(): Promise<boolean> {
   // Then ensure the Ministral 3 8B model is pulled
   return ensureOllamaModel('ministral-3:8b');
 }
+
+/**
+ * Ensure Qwen3-VL 8B model is available (vision-language model)
+ * Q4_K_M quantization (~5GB) - fits in 15GB VRAM with room to spare
+ */
+export async function ensureQwen3Vl(): Promise<boolean> {
+  // First ensure the Ollama service is running
+  const ollamaOk = await ensureMiniCpm();
+  if (!ollamaOk) return false;
+
+  // Then ensure Qwen3-VL 8B is pulled
+  return ensureOllamaModel('qwen3-vl:8b');
+}

test/test.invoices.ministral3.ts

@@ -36,8 +36,9 @@ function convertPdfToImages(pdfPath: string): string[] {
   const outputPattern = path.join(tempDir, 'page-%d.png');
 
   try {
+    // High quality conversion: 300 DPI, max quality, sharpen for better OCR
     execSync(
-      `convert -density 200 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
+      `convert -density 300 -quality 100 "${pdfPath}" -background white -alpha remove -sharpen 0x1 "${outputPattern}"`,
       { stdio: 'pipe' }
     );
 
@@ -77,18 +78,35 @@ async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
     required: ['invoice_number', 'invoice_date', 'vendor_name', 'currency', 'net_amount', 'vat_amount', 'total_amount'],
   };
 
-  const prompt = `Extract invoice data from this document image(s).
+  const prompt = `You are an expert invoice data extraction system. Carefully analyze this invoice document and extract the following fields with high precision.
 
-Find and return:
-- invoice_number: The invoice number/ID (look for "Invoice No", "Invoice #", "Rechnung Nr")
-- invoice_date: The invoice date in YYYY-MM-DD format
-- vendor_name: The company issuing the invoice (in letterhead)
-- currency: EUR, USD, or GBP
-- total_amount: The FINAL total amount due
-- net_amount: Amount before VAT/tax
-- vat_amount: VAT/tax amount
+INVOICE NUMBER:
+- Look for labels: "Invoice No", "Invoice #", "Invoice Number", "Rechnung Nr", "Rechnungsnummer", "Document No", "Bill No", "Reference"
+- Usually alphanumeric, often starts with letters (e.g., R0014359508, INV-2024-001)
+- Located near the top of the invoice
 
-Return ONLY valid JSON.`;
+INVOICE DATE:
+- Look for labels: "Invoice Date", "Date", "Datum", "Rechnungsdatum", "Issue Date", "Bill Date"
+- Convert ANY date format to YYYY-MM-DD (e.g., 14/10/2021 → 2021-10-14, Oct 14, 2021 → 2021-10-14)
+- Usually near the invoice number
+
+VENDOR NAME:
+- The company ISSUING the invoice (not the recipient)
+- Found in letterhead, logo area, or header - typically the largest/most prominent company name
+- Examples: "Hetzner Online GmbH", "Adobe Inc", "DigitalOcean LLC"
+
+CURRENCY:
+- Detect from symbols: € = EUR, $ = USD, £ = GBP
+- Or from text: "EUR", "USD", "GBP"
+- Default to EUR if unclear
+
+AMOUNTS (Critical - read carefully!):
+- total_amount: The FINAL amount due/payable - look for "Total", "Grand Total", "Amount Due", "Balance Due", "Gesamtbetrag", "Endbetrag"
+- net_amount: Subtotal BEFORE tax - look for "Subtotal", "Net", "Netto", "excl. VAT"
+- vat_amount: Tax amount - look for "VAT", "Tax", "MwSt", "USt", "19%", "20%"
+- For multi-page invoices: the FINAL totals are usually on the LAST page
+
+Return ONLY valid JSON with the extracted values.`;
 
   const response = await fetch(`${OLLAMA_URL}/api/chat`, {
     method: 'POST',
@@ -105,7 +123,7 @@ Return ONLY valid JSON.`;
       format: invoiceSchema,
       stream: true,
       options: {
-        num_predict: 512,
+        num_predict: 1024,
         temperature: 0.0,
       },
     }),
@@ -170,46 +188,6 @@ Return ONLY valid JSON.`;
   };
 }
 
-/**
- * Extract with consensus voting (2 agreeing passes)
- */
-async function extractWithConsensus(images: string[], name: string, maxPasses: number = 3): Promise<IInvoice> {
-  const results: Array<{ invoice: IInvoice; hash: string }> = [];
-  const hashCounts: Map<string, number> = new Map();
-
-  for (let pass = 1; pass <= maxPasses; pass++) {
-    try {
-      const invoice = await extractInvoiceFromImages(images);
-      const hash = `${invoice.invoice_number}|${invoice.invoice_date}|${invoice.total_amount?.toFixed(2)}`;
-      results.push({ invoice, hash });
-      hashCounts.set(hash, (hashCounts.get(hash) || 0) + 1);
-
-      console.log(`  [Pass ${pass}] ${invoice.invoice_number} | ${invoice.invoice_date} | ${invoice.total_amount} ${invoice.currency}`);
-
-      if (hashCounts.get(hash)! >= 2) {
-        console.log(`  [Consensus] Reached after ${pass} passes`);
-        return invoice;
-      }
-    } catch (err) {
-      console.log(`  [Pass ${pass}] Error: ${err}`);
-    }
-  }
-
-  // Return most common result
-  let bestHash = '';
-  let bestCount = 0;
-  for (const [hash, count] of hashCounts) {
-    if (count > bestCount) {
-      bestCount = count;
-      bestHash = hash;
-    }
-  }
-
-  if (!bestHash) throw new Error(`No valid results for ${name}`);
-
-  console.log(`  [No consensus] Using best result (${bestCount}/${maxPasses})`);
-  return results.find((r) => r.hash === bestHash)!.invoice;
-}
 
 /**
  * Normalize date to YYYY-MM-DD
@@ -314,7 +292,8 @@ for (const testCase of testCases) {
     const images = convertPdfToImages(testCase.pdfPath);
     console.log(`  Pages: ${images.length}`);
 
-    const extracted = await extractWithConsensus(images, testCase.name);
+    const extracted = await extractInvoiceFromImages(images);
+    console.log(`  Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
     const elapsed = Date.now() - start;
     times.push(elapsed);
 

test/test.invoices.paddleocr-vl.ts

@@ -89,25 +89,13 @@ async function parseDocument(imageBase64: string): Promise<string> {
   return data.result?.html || '';
 }
 
-/**
- * Sanitize HTML to remove OCR artifacts that confuse the LLM
- * Minimal cleaning - only remove truly problematic patterns
- */
-function sanitizeHtml(html: string): string {
-  // Remove excessively repeated characters (OCR glitches)
-  let sanitized = html.replace(/(\d)\1{20,}/g, '$1...');
-  // Remove extremely long strings (corrupted data)
-  sanitized = sanitized.replace(/\b[A-Za-z0-9]{50,}\b/g, '[OCR_ARTIFACT]');
-  return sanitized;
-}
-
 /**
  * Extract invoice fields using simple direct prompt
  * The OCR output has clearly labeled fields - just ask the LLM to read them
  */
 async function extractInvoiceFromHtml(html: string): Promise<IInvoice> {
-  const sanitized = sanitizeHtml(html);
-  const truncated = sanitized.length > 32000 ? sanitized.slice(0, 32000) : sanitized;
+  // OCR output is already good - just truncate if too long
+  const truncated = html.length > 32000 ? html.slice(0, 32000) : html;
   console.log(`    [Extract] ${truncated.length} chars of HTML`);
 
   // JSON schema for structured output

test/test.invoices.qwen3vl.ts

@@ -0,0 +1,311 @@
+/**
+ * Invoice extraction using Qwen3-VL 8B Vision (Direct)
+ *
+ * Single-step pipeline: PDF → Images → Qwen3-VL → JSON
+ * Uses /no_think to disable reasoning mode for fast, direct responses.
+ *
+ * Qwen3-VL outperforms PaddleOCR-VL on certain invoice formats.
+ */
+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 { ensureMiniCpm } from './helpers/docker.js';
+
+const OLLAMA_URL = 'http://localhost:11434';
+const VISION_MODEL = 'qwen3-vl:8b';
+
+interface IInvoice {
+  invoice_number: string;
+  invoice_date: string;
+  vendor_name: string;
+  currency: string;
+  net_amount: number;
+  vat_amount: number;
+  total_amount: number;
+}
+
+/**
+ * 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 {
+    // 150 DPI is sufficient for invoice extraction, reduces context size
+    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 });
+  }
+}
+
+/**
+ * Extract invoice data directly from images using Qwen3-VL Vision
+ * Uses /no_think to disable reasoning mode for fast, direct JSON output
+ */
+async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
+  console.log(`    [Vision] Processing ${images.length} page(s) with Qwen3-VL`);
+
+  // /no_think disables Qwen3's reasoning mode - crucial for getting direct output
+  const prompt = `/no_think
+Look at this invoice and extract these fields. Reply with ONLY JSON, no explanation.
+
+- invoice_number
+- invoice_date (format: YYYY-MM-DD)
+- vendor_name
+- currency (EUR, USD, or GBP)
+- net_amount
+- vat_amount
+- total_amount
+
+JSON: {"invoice_number":"...","invoice_date":"YYYY-MM-DD","vendor_name":"...","currency":"EUR","net_amount":0,"vat_amount":0,"total_amount":0}`;
+
+  const response = await fetch(`${OLLAMA_URL}/api/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: VISION_MODEL,
+      messages: [{
+        role: 'user',
+        content: prompt,
+        images: images, // Pass all pages
+      }],
+      stream: false,
+      options: {
+        num_predict: 512,
+        temperature: 0.0,
+      },
+    }),
+  });
+
+  if (!response.ok) {
+    const err = await response.text();
+    throw new Error(`Ollama API error: ${response.status} - ${err}`);
+  }
+
+  const data = await response.json();
+  let content = data.message?.content || '';
+
+  console.log(`    [Vision] Response (${content.length} chars): ${content.substring(0, 200)}...`);
+
+  // Parse JSON from response
+  if (content.startsWith('```json')) content = content.slice(7);
+  else if (content.startsWith('```')) content = content.slice(3);
+  if (content.endsWith('```')) content = content.slice(0, -3);
+  content = content.trim();
+
+  const startIdx = content.indexOf('{');
+  const endIdx = content.lastIndexOf('}') + 1;
+
+  if (startIdx < 0 || endIdx <= startIdx) {
+    throw new Error(`No JSON found: ${content.substring(0, 300)}`);
+  }
+
+  const parsed = JSON.parse(content.substring(startIdx, endIdx));
+
+  return {
+    invoice_number: parsed.invoice_number || null,
+    invoice_date: parsed.invoice_date || null,
+    vendor_name: parsed.vendor_name || null,
+    currency: parsed.currency || 'EUR',
+    net_amount: parseFloat(parsed.net_amount) || 0,
+    vat_amount: parseFloat(parsed.vat_amount) || 0,
+    total_amount: parseFloat(parsed.total_amount) || 0,
+  };
+}
+
+/**
+ * Normalize date to YYYY-MM-DD
+ */
+function normalizeDate(dateStr: string | null): string {
+  if (!dateStr) return '';
+  if (/^\d{4}-\d{2}-\d{2}$/.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 vs 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: expected "${expected.invoice_number}", got "${extracted.invoice_number}"`);
+  }
+
+  if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
+    errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
+  }
+
+  if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
+    errors.push(`total_amount: expected ${expected.total_amount}, got ${extracted.total_amount}`);
+  }
+
+  if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
+    errors.push(`currency: expected "${expected.currency}", got "${extracted.currency}"`);
+  }
+
+  return { match: errors.length === 0, errors };
+}
+
+/**
+ * Find test cases
+ */
+function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
+  const testDir = path.join(process.cwd(), '.nogit/invoices');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
+
+  for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${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));
+}
+
+/**
+ * Ensure Qwen3-VL 8B model is available
+ */
+async function ensureQwen3Vl(): 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 === VISION_MODEL)) {
+        console.log(`[Ollama] Model already available: ${VISION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    console.log('[Ollama] Cannot check models');
+    return false;
+  }
+
+  console.log(`[Ollama] Pulling model: ${VISION_MODEL}...`);
+  const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ name: VISION_MODEL, stream: false }),
+  });
+
+  return pullResponse.ok;
+}
+
+// Tests
+
+tap.test('setup: ensure Qwen3-VL is running', async () => {
+  console.log('\n[Setup] Checking Qwen3-VL 8B...\n');
+
+  // Ensure Ollama service is running
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+
+  // Ensure Qwen3-VL 8B model
+  const visionOk = await ensureQwen3Vl();
+  expect(visionOk).toBeTrue();
+
+  console.log('\n[Setup] Ready!\n');
+});
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} invoice test cases (Qwen3-VL Vision)\n`);
+
+let passedCount = 0;
+let failedCount = 0;
+const times: number[] = [];
+
+for (const testCase of testCases) {
+  tap.test(`should extract invoice: ${testCase.name}`, async () => {
+    const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
+    console.log(`\n=== ${testCase.name} ===`);
+    console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
+
+    const start = Date.now();
+    const images = convertPdfToImages(testCase.pdfPath);
+    console.log(`  Pages: ${images.length}`);
+
+    const extracted = await extractInvoiceFromImages(images);
+    console.log(`  Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
+    const elapsed = Date.now() - start;
+    times.push(elapsed);
+
+    const result = compareInvoice(extracted, expected);
+
+    if (result.match) {
+      passedCount++;
+      console.log(`  Result: MATCH (${(elapsed / 1000).toFixed(1)}s)`);
+    } else {
+      failedCount++;
+      console.log(`  Result: MISMATCH (${(elapsed / 1000).toFixed(1)}s)`);
+      result.errors.forEach((e) => console.log(`    - ${e}`));
+    }
+
+    expect(result.match).toBeTrue();
+  });
+}
+
+tap.test('summary', async () => {
+  const total = testCases.length;
+  const accuracy = total > 0 ? (passedCount / total) * 100 : 0;
+  const totalTime = times.reduce((a, b) => a + b, 0) / 1000;
+  const avgTime = times.length > 0 ? totalTime / times.length : 0;
+
+  console.log(`\n======================================================`);
+  console.log(`   Invoice Extraction Summary (Qwen3-VL Vision)`);
+  console.log(`======================================================`);
+  console.log(`  Method:    Qwen3-VL 8B Direct Vision (/no_think)`);
+  console.log(`  Passed:    ${passedCount}/${total}`);
+  console.log(`  Failed:    ${failedCount}/${total}`);
+  console.log(`  Accuracy:  ${accuracy.toFixed(1)}%`);
+  console.log(`------------------------------------------------------`);
+  console.log(`  Total time:   ${totalTime.toFixed(1)}s`);
+  console.log(`  Avg per inv:  ${avgTime.toFixed(1)}s`);
+  console.log(`======================================================\n`);
+});
+
+export default tap.start();