feat(tests): switch vision tests to multi-query extraction (count then per-row/field queries) and add logging/summaries

This commit is contained in:
2026-01-18 11:26:38 +00:00
parent 4c368dfef9
commit 76b21f1f7b
5 changed files with 624 additions and 367 deletions

View File

@@ -1,12 +1,10 @@
/**
* 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
* Multi-query approach:
* 1. First ask how many transactions on each page
* 2. Then query each transaction individually
* Single pass, no consensus voting.
*/
import { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs';
@@ -53,15 +51,9 @@ function convertPdfToImages(pdfPath: string): string[] {
}
/**
* Extract transactions from a single page
* Processes one page at a time to minimize thinking tokens
* Query Qwen3-VL with a simple prompt
*/
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},...]`;
async function queryVision(image: string, prompt: string): Promise<string> {
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
@@ -73,9 +65,8 @@ Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21
images: [image],
}],
stream: false,
think: false,
options: {
num_predict: 4000,
num_predict: 500,
temperature: 0.1,
},
}),
@@ -86,47 +77,116 @@ Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21
}
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 [];
}
return (data.message?.content || '').trim();
}
/**
* Extract transactions using Qwen3-VL vision
* Processes each page separately to avoid thinking token exhaustion
* Count transactions on a page
*/
async function countTransactions(image: string, pageNum: number): Promise<number> {
const response = await queryVision(image,
`How many transaction rows are in this bank statement table?
Count only the data rows (with dates like "01.01.2024" and amounts like "- 50,00 €").
Do NOT count the header row or summary/total rows.
Answer with just the number, for example: 7`
);
console.log(` [Page ${pageNum}] Count query response: "${response}"`);
const match = response.match(/(\d+)/);
const count = match ? parseInt(match[1], 10) : 0;
console.log(` [Page ${pageNum}] Parsed count: ${count}`);
return count;
}
/**
* Get a single transaction by index (logs immediately when complete)
*/
async function getTransaction(image: string, index: number, pageNum: number): Promise<ITransaction | null> {
const response = await queryVision(image,
`This is a bank statement. Look at transaction row #${index} in the table (counting from top, excluding headers).
Extract this transaction's details:
- Date in YYYY-MM-DD format
- Counterparty/description name
- Amount as number (negative for debits like "- 21,47 €" = -21.47, positive for credits like "+ 100,00 €" = 100.00)
Answer in format: DATE|COUNTERPARTY|AMOUNT
Example: 2024-01-15|Amazon|25.99`
);
// Parse the response
const lines = response.split('\n').filter(l => l.includes('|'));
const line = lines[lines.length - 1] || response;
const parts = line.split('|').map(p => p.trim());
if (parts.length >= 3) {
// Parse amount - handle various formats
let amountStr = parts[2].replace(/[€$£\s]/g, '').replace('', '-').replace('', '-');
// European format: comma is decimal
if (amountStr.includes(',')) {
amountStr = amountStr.replace(/\./g, '').replace(',', '.');
}
const amount = parseFloat(amountStr) || 0;
const tx = {
date: parts[0],
counterparty: parts[1],
amount: amount,
};
// Log immediately as this transaction completes
console.log(` [P${pageNum} Tx${index.toString().padStart(2, ' ')}] ${tx.date} | ${tx.counterparty.substring(0, 25).padEnd(25)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
return tx;
}
// Log raw response on parse failure
console.log(` [P${pageNum} Tx${index.toString().padStart(2, ' ')}] PARSE FAILED: "${response.replace(/\n/g, ' ').substring(0, 60)}..."`);
return null;
}
/**
* Extract transactions from a single page using multi-query approach
*/
async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
// Step 1: Count transactions
const count = await countTransactions(image, pageNum);
if (count === 0) {
return [];
}
// Step 2: Query each transaction (in batches to avoid overwhelming)
// Each transaction logs itself as it completes
const transactions: ITransaction[] = [];
const batchSize = 5;
for (let start = 1; start <= count; start += batchSize) {
const end = Math.min(start + batchSize - 1, count);
const indices = Array.from({ length: end - start + 1 }, (_, i) => start + i);
// Query batch in parallel - each logs as it completes
const results = await Promise.all(
indices.map(i => getTransaction(image, i, pageNum))
);
for (const tx of results) {
if (tx) {
transactions.push(tx);
}
}
}
console.log(` [Page ${pageNum}] Complete: ${transactions.length}/${count} extracted`);
return transactions;
}
/**
* Extract all transactions from bank statement
*/
async function extractTransactions(images: string[]): Promise<ITransaction[]> {
console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL`);
console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`);
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);
@@ -276,8 +336,9 @@ tap.test('summary', async () => {
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(` Method: Multi-query (count then extract each)`);
console.log(` Passed: ${passedCount}/${total}`);
console.log(` Failed: ${failedCount}/${total}`);
console.log(`======================================================\n`);
});