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

changelog.md

@@ -1,5 +1,15 @@
 # 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
 

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';
@@ -57,144 +56,68 @@ function convertPdfToImages(pdfPath: string): string[] {
 }
 
 /**
- * Single extraction attempt
+ * Extract invoice data directly from images using Qwen3-VL Vision
+ * Uses /no_think to disable reasoning mode for fast, direct JSON output
  */
-async function tryExtractOnce(images: string[], prompt: string): Promise<string> {
+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: [
+      messages: [{
-        {
         role: 'user',
         content: prompt,
-          images: images,
+        images: images, // Pass all pages
-        },
+      }],
-      ],
+      stream: false,
-      stream: true,
       options: {
-        num_predict: 1024,
+        num_predict: 512,
-        temperature: 0.1, // Slight randomness helps avoid stuck states
+        temperature: 0.0,
       },
     }),
   });
 
   if (!response.ok) {
-    throw new Error(`Ollama API error: ${response.status}`);
+    const err = await response.text();
+    throw new Error(`Ollama API error: ${response.status} - ${err}`);
   }
 
-  const reader = response.body?.getReader();
+  const data = await response.json();
-  if (!reader) {
+  let content = data.message?.content || '';
-    throw new Error('No response body');
-  }
 
-  const decoder = new TextDecoder();
+  console.log(`    [Vision] Response (${content.length} chars): ${content.substring(0, 200)}...`);
-  let fullText = '';
 
-  while (true) {
+  // Parse JSON from response
-    const { done, value } = await reader.read();
+  if (content.startsWith('```json')) content = content.slice(7);
-    if (done) break;
+  else if (content.startsWith('```')) content = content.slice(3);
+  if (content.endsWith('```')) content = content.slice(0, -3);
+  content = content.trim();
 
-    const chunk = decoder.decode(value, { stream: true });
+  const startIdx = content.indexOf('{');
-    const lines = chunk.split('\n').filter((l) => l.trim());
+  const endIdx = content.lastIndexOf('}') + 1;
-
-    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
- * Includes retry logic for empty responses
- */
-async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
-  console.log(`    [Vision] Processing ${images.length} page(s) with Qwen3-VL`);
-
-  // JSON schema for structured output - force the model to output valid JSON
-  const invoiceSchema = {
-    type: 'object',
-    properties: {
-      invoice_number: { type: 'string' },
-      invoice_date: { type: 'string' },
-      vendor_name: { type: 'string' },
-      currency: { type: 'string' },
-      net_amount: { type: 'number' },
-      vat_amount: { type: 'number' },
-      total_amount: { type: 'number' },
-    },
-    required: ['invoice_number', 'invoice_date', 'vendor_name', 'currency', 'net_amount', 'vat_amount', 'total_amount'],
-  };
-
-  // 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)}`);
+    throw new Error(`No JSON found: ${content.substring(0, 300)}`);
   }
 
-  const extractedJson = jsonStr.substring(startIdx, endIdx);
+  const parsed = JSON.parse(content.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,
@@ -284,12 +207,48 @@ 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 +298,7 @@ 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)}%`);