v1.11.0

changelog.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## 2026-01-18 - 1.11.0 - feat(vision)
+process pages separately and make Qwen3-VL vision extraction more robust; add per-page parsing, safer JSON handling, reduced token usage, and multi-query invoice extraction
+
+- Bank statements: split extraction into extractTransactionsFromPage and sequentially process pages to avoid thinking-token exhaustion
+- Bank statements: reduced num_predict from 8000 to 4000, send single image per request, added per-page logging and non-throwing handling for empty or non-JSON responses
+- Bank statements: catch JSON.parse errors and return empty array instead of throwing
+- Invoices: introduced queryField to request single values and perform multiple simple queries (reduces model thinking usage)
+- Invoices: reduced num_predict for invoice queries from 4000 to 500 and parse amounts robustly (handles European formats like 1.234,56)
+- Invoices: normalize currency to uppercase 3-letter code, return safe defaults (empty strings / 0) instead of nulls, and parse net/vat/total with fallbacks
+- General: simplified Ollama API error messages to avoid including response body content in thrown errors
+
+## 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
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@host.today/ht-docker-ai",
-  "version": "1.10.0",
+  "version": "1.11.0",
   "type": "module",
   "private": false,
   "description": "Docker images for AI vision-language models including MiniCPM-V 4.5",

test/test.bankstatements.qwen3vl.ts

@@ -0,0 +1,284 @@
+/**
+ * Bank statement extraction using Qwen3-VL 8B Vision (Direct)
+ *
+ * Single-step pipeline: PDF → Images → Qwen3-VL → JSON
+ *
+ * Key insights:
+ * - Use /no_think in prompt + think:false in API to disable reasoning
+ * - Need high num_predict (8000+) for many transactions
+ * - Single pass extraction, no consensus needed
+ */
+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 ITransaction {
+  date: string;
+  counterparty: string;
+  amount: number;
+}
+
+/**
+ * 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: 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 });
+  }
+}
+
+/**
+ * Extract transactions from a single page
+ * Processes one page at a time to minimize thinking tokens
+ */
+async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
+  const prompt = `/no_think
+Extract transactions from this bank statement page.
+Amount: "- 21,47 €" = -21.47, "+ 1.000,00 €" = 1000.00 (European format)
+Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21.47},...]`;
+
+  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: [image],
+      }],
+      stream: false,
+      think: false,
+      options: {
+        num_predict: 4000,
+        temperature: 0.1,
+      },
+    }),
+  });
+
+  if (!response.ok) {
+    throw new Error(`Ollama API error: ${response.status}`);
+  }
+
+  const data = await response.json();
+  let content = data.message?.content || '';
+
+  if (!content) {
+    console.log(`    [Page ${pageNum}] Empty response`);
+    return [];
+  }
+
+  // Parse JSON array
+  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) {
+    console.log(`    [Page ${pageNum}] No JSON array found`);
+    return [];
+  }
+
+  try {
+    const transactions = JSON.parse(content.substring(startIdx, endIdx));
+    console.log(`    [Page ${pageNum}] Found ${transactions.length} transactions`);
+    return transactions;
+  } catch {
+    console.log(`    [Page ${pageNum}] JSON parse error`);
+    return [];
+  }
+}
+
+/**
+ * Extract transactions using Qwen3-VL vision
+ * Processes each page separately to avoid thinking token exhaustion
+ */
+async function extractTransactions(images: string[]): Promise<ITransaction[]> {
+  console.log(`    [Vision] Processing ${images.length} page(s) with Qwen3-VL`);
+
+  const allTransactions: ITransaction[] = [];
+
+  // Process pages sequentially to avoid overwhelming the model
+  for (let i = 0; i < images.length; i++) {
+    const pageTransactions = await extractTransactionsFromPage(images[i], i + 1);
+    allTransactions.push(...pageTransactions);
+  }
+
+  console.log(`    [Vision] Total: ${allTransactions.length} transactions`);
+  return allTransactions;
+}
+
+/**
+ * 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 transaction ${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 at ${i}: expected ${exp.date}/${exp.amount}, got ${ext.date}/${ext.amount}`);
+    }
+  }
+
+  if (extracted.length > expected.length) {
+    errors.push(`Extra transactions: ${extracted.length - expected.length}`);
+  }
+
+  return { matches, total: expected.length, errors };
+}
+
+/**
+ * Find test cases in .nogit/
+ */
+function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
+  const testDir = path.join(process.cwd(), '.nogit');
+  if (!fs.existsSync(testDir)) return [];
+
+  const files = fs.readdirSync(testDir);
+  const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
+
+  for (const pdf of files.filter((f: string) => f.endsWith('.pdf'))) {
+    const baseName = pdf.replace('.pdf', '');
+    const jsonFile = `${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));
+}
+
+/**
+ * Ensure Qwen3-VL 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 available: ${VISION_MODEL}`);
+        return true;
+      }
+    }
+  } catch {
+    return false;
+  }
+
+  console.log(`[Ollama] Pulling ${VISION_MODEL}...`);
+  const pullResponse = await fetch(`${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');
+  const ollamaOk = await ensureMiniCpm();
+  expect(ollamaOk).toBeTrue();
+  const visionOk = await ensureQwen3Vl();
+  expect(visionOk).toBeTrue();
+  console.log('\n[Setup] Ready!\n');
+});
+
+const testCases = findTestCases();
+console.log(`\nFound ${testCases.length} bank statement test cases (Qwen3-VL)\n`);
+
+let passedCount = 0;
+let failedCount = 0;
+
+for (const testCase of testCases) {
+  tap.test(`should extract: ${testCase.name}`, async () => {
+    const expected: ITransaction[] = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
+    console.log(`\n=== ${testCase.name} ===`);
+    console.log(`Expected: ${expected.length} transactions`);
+
+    const images = convertPdfToImages(testCase.pdfPath);
+    console.log(`  Pages: ${images.length}`);
+
+    const extracted = await extractTransactions(images);
+    console.log(`  Extracted: ${extracted.length} transactions`);
+
+    const result = compareTransactions(extracted, expected);
+    const accuracy = result.total > 0 ? result.matches / result.total : 0;
+
+    if (accuracy >= 0.95 && extracted.length === expected.length) {
+      passedCount++;
+      console.log(`  Result: PASS (${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(accuracy).toBeGreaterThan(0.95);
+    expect(extracted.length).toEqual(expected.length);
+  });
+}
+
+tap.test('summary', async () => {
+  const total = testCases.length;
+  console.log(`\n======================================================`);
+  console.log(`   Bank Statement Summary (Qwen3-VL Vision)`);
+  console.log(`======================================================`);
+  console.log(`  Passed: ${passedCount}/${total}`);
+  console.log(`  Failed: ${failedCount}/${total}`);
+  console.log(`======================================================\n`);
+});
+
+export default tap.start();

test/test.invoices.qwen3vl.ts

@@ -1,18 +1,17 @@
 /**
- * Invoice extraction using Qwen3-VL-8B Vision (Direct)
+ * Invoice extraction using Qwen3-VL 8B Vision (Direct)
  *
- * Qwen3-VL 8B is a capable vision-language model that fits in 15GB VRAM:
+ * Single-step pipeline: PDF → Images → Qwen3-VL → JSON
- *   - Q4_K_M quantization (~5GB)
+ * Uses /no_think to disable reasoning mode for fast, direct responses.
- *   - Good balance of speed and accuracy
  *
- * Pipeline: PDF → Images → Qwen3-VL → JSON
+ * 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 { ensureQwen3Vl } from './helpers/docker.js';
+import { ensureMiniCpm } from './helpers/docker.js';
 
 const OLLAMA_URL = 'http://localhost:11434';
 const VISION_MODEL = 'qwen3-vl:8b';
@@ -57,25 +56,25 @@ function convertPdfToImages(pdfPath: string): string[] {
 }
 
 /**
- * Single extraction attempt
+ * Query Qwen3-VL for a single field
+ * Uses simple prompts to minimize thinking tokens
  */
-async function tryExtractOnce(images: string[], prompt: string): Promise<string> {
+async function queryField(images: string[], question: string): Promise<string> {
   const response = await fetch(`${OLLAMA_URL}/api/chat`, {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
     body: JSON.stringify({
       model: VISION_MODEL,
-      messages: [
+      messages: [{
-        {
+        role: 'user',
-          role: 'user',
+        content: `/no_think\n${question} Reply with just the value, nothing else.`,
-          content: prompt,
+        images: images,
-          images: images,
+      }],
-        },
+      stream: false,
-      ],
+      think: false,
-      stream: true,
       options: {
-        num_predict: 1024,
+        num_predict: 500,
-        temperature: 0.1, // Slight randomness helps avoid stuck states
+        temperature: 0.1,
       },
     }),
   });
@@ -84,126 +83,48 @@ async function tryExtractOnce(images: string[], prompt: string): Promise<string>
     throw new Error(`Ollama API error: ${response.status}`);
   }
 
-  const reader = response.body?.getReader();
+  const data = await response.json();
-  if (!reader) {
+  return (data.message?.content || '').trim();
-    throw new Error('No response body');
-  }
-
-  const decoder = new TextDecoder();
-  let fullText = '';
-
-  while (true) {
-    const { done, value } = await reader.read();
-    if (done) break;
-
-    const chunk = decoder.decode(value, { stream: true });
-    const lines = chunk.split('\n').filter((l) => l.trim());
-
-    for (const line of lines) {
-      try {
-        const json = JSON.parse(line);
-        if (json.message?.content) {
-          fullText += json.message.content;
-        }
-      } catch {
-        // Skip invalid JSON lines
-      }
-    }
-  }
-
-  return fullText;
 }
 
 /**
- * Extract invoice data directly from images using Qwen3-VL Vision
+ * Extract invoice data using multiple simple queries
- * Includes retry logic for empty responses
+ * Each query asks for 1-2 fields to minimize thinking tokens
+ * (Qwen3's thinking mode uses all tokens on complex prompts)
  */
 async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
-  console.log(`    [Vision] Processing ${images.length} page(s) with Qwen3-VL`);
+  console.log(`    [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`);
 
-  // JSON schema for structured output - force the model to output valid JSON
+  // Query each field separately to avoid excessive thinking tokens
-  const invoiceSchema = {
+  const [invoiceNum, invoiceDate, vendor, currency, amounts] = await Promise.all([
-    type: 'object',
+    queryField(images, 'What is the invoice number on this document?'),
-    properties: {
+    queryField(images, 'What is the invoice date? Format as YYYY-MM-DD.'),
-      invoice_number: { type: 'string' },
+    queryField(images, 'What company issued this invoice?'),
-      invoice_date: { type: 'string' },
+    queryField(images, 'What currency is used? Answer EUR, USD, or GBP.'),
-      vendor_name: { type: 'string' },
+    queryField(images, 'What are the net amount, VAT amount, and total amount? Format: net,vat,total'),
-      currency: { type: 'string' },
+  ]);
-      net_amount: { type: 'number' },
+
-      vat_amount: { type: 'number' },
+  console.log(`    [Vision] Got: ${invoiceNum} | ${invoiceDate} | ${vendor} | ${currency}`);
-      total_amount: { type: 'number' },
+
-    },
+  // Parse amounts (format: "net,vat,total" or similar)
-    required: ['invoice_number', 'invoice_date', 'vendor_name', 'currency', 'net_amount', 'vat_amount', 'total_amount'],
+  const amountMatch = amounts.match(/([\d.,]+)/g) || [];
+  const parseAmount = (s: string): number => {
+    if (!s) return 0;
+    // Handle European format: 1.234,56 → 1234.56
+    const normalized = s.includes(',') && s.indexOf(',') > s.lastIndexOf('.')
+      ? s.replace(/\./g, '').replace(',', '.')
+      : s.replace(/,/g, '');
+    return parseFloat(normalized) || 0;
   };
 
-  // Simple, direct prompt - don't overthink, just read the labeled fields
-  const prompt = `Extract invoice data from this image. Return JSON only.
-
-Find these fields:
-- invoice_number: The invoice/document number
-- invoice_date: Date in YYYY-MM-DD format
-- vendor_name: Company issuing the invoice
-- currency: EUR, USD, or GBP
-- net_amount: Amount before tax
-- vat_amount: Tax/VAT amount
-- total_amount: Final total amount
-
-Return: {"invoice_number":"...", "invoice_date":"YYYY-MM-DD", "vendor_name":"...", "currency":"EUR", "net_amount":0.00, "vat_amount":0.00, "total_amount":0.00}`;
-
-  // Retry logic for empty responses (model sometimes returns nothing)
-  const MAX_RETRIES = 3;
-  let fullText = '';
-
-  for (let attempt = 1; attempt <= MAX_RETRIES; attempt++) {
-    fullText = await tryExtractOnce(images, prompt);
-
-    if (fullText.trim().length > 0) {
-      console.log(`    [Attempt ${attempt}] Got ${fullText.length} chars`);
-      break;
-    }
-
-    console.log(`    [Attempt ${attempt}] Empty response, retrying...`);
-    // Small delay before retry
-    await new Promise((r) => setTimeout(r, 1000));
-  }
-
-  if (fullText.trim().length === 0) {
-    throw new Error(`Model returned empty response after ${MAX_RETRIES} attempts`);
-  }
-
-  // Parse JSON response
-  let jsonStr = fullText.trim();
-
-  if (jsonStr.startsWith('```json')) jsonStr = jsonStr.slice(7);
-  else if (jsonStr.startsWith('```')) jsonStr = jsonStr.slice(3);
-  if (jsonStr.endsWith('```')) jsonStr = jsonStr.slice(0, -3);
-  jsonStr = jsonStr.trim();
-
-  const startIdx = jsonStr.indexOf('{');
-  const endIdx = jsonStr.lastIndexOf('}') + 1;
-
-  if (startIdx < 0 || endIdx <= startIdx) {
-    throw new Error(`No JSON found in: ${fullText.substring(0, 500)}`);
-  }
-
-  const extractedJson = jsonStr.substring(startIdx, endIdx);
-  console.log(`    [Debug] Extracted JSON: ${extractedJson.substring(0, 200)}...`);
-
-  let parsed;
-  try {
-    parsed = JSON.parse(extractedJson);
-  } catch (e) {
-    throw new Error(`Invalid JSON: ${extractedJson.substring(0, 500)}`);
-  }
-
   return {
-    invoice_number: parsed.invoice_number || null,
+    invoice_number: invoiceNum || '',
-    invoice_date: parsed.invoice_date || null,
+    invoice_date: invoiceDate || '',
-    vendor_name: parsed.vendor_name || null,
+    vendor_name: vendor || '',
-    currency: parsed.currency || 'EUR',
+    currency: (currency || 'EUR').toUpperCase().replace(/[^A-Z]/g, '').slice(0, 3) || 'EUR',
-    net_amount: parseFloat(parsed.net_amount) || 0,
+    net_amount: parseAmount(amountMatch[0] || ''),
-    vat_amount: parseFloat(parsed.vat_amount) || 0,
+    vat_amount: parseAmount(amountMatch[1] || ''),
-    total_amount: parseFloat(parsed.total_amount) || 0,
+    total_amount: parseAmount(amountMatch[2] || amountMatch[0] || ''),
   };
 }
 
@@ -284,12 +205,48 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
   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 (~5GB)...\n');
+  console.log('\n[Setup] Checking Qwen3-VL 8B...\n');
-  const ok = await ensureQwen3Vl();
+
-  expect(ok).toBeTrue();
+  // 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');
 });
 
@@ -339,7 +296,7 @@ tap.test('summary', async () => {
   console.log(`\n======================================================`);
   console.log(`   Invoice Extraction Summary (Qwen3-VL Vision)`);
   console.log(`======================================================`);
-  console.log(`  Method:    Qwen3-VL 8B (Direct Vision)`);
+  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)}%`);