feat(tests): switch vision tests to multi-query extraction (count then per-row/field queries) and add logging/summaries
This commit is contained in:
@@ -1,8 +1,10 @@
|
||||
/**
|
||||
* Bank statement extraction test using MiniCPM-V only (visual extraction)
|
||||
* Bank statement extraction using MiniCPM-V (visual extraction)
|
||||
*
|
||||
* This tests MiniCPM-V's ability to extract bank transactions directly from images
|
||||
* without any OCR augmentation.
|
||||
* Multi-query approach with thinking DISABLED for speed:
|
||||
* 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';
|
||||
@@ -11,24 +13,8 @@ import { execSync } from 'child_process';
|
||||
import * as os from 'os';
|
||||
import { ensureMiniCpm } from './helpers/docker.js';
|
||||
|
||||
// Service URL
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
|
||||
// Model
|
||||
const MINICPM_MODEL = 'minicpm-v:latest';
|
||||
|
||||
// Prompt for MiniCPM-V visual extraction
|
||||
const MINICPM_EXTRACT_PROMPT = `/nothink
|
||||
You are a bank statement parser. Extract EVERY transaction from the table.
|
||||
|
||||
Read the Amount column carefully:
|
||||
- "- 21,47 €" means DEBIT, output as: -21.47
|
||||
- "+ 1.000,00 €" means CREDIT, output as: 1000.00
|
||||
- European format: comma = decimal point
|
||||
|
||||
For each row output: {"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21.47}
|
||||
|
||||
Do not skip any rows. Return ONLY the JSON array, no explanation.`;
|
||||
const MODEL = 'minicpm-v:latest';
|
||||
|
||||
interface ITransaction {
|
||||
date: string;
|
||||
@@ -65,149 +51,146 @@ function convertPdfToImages(pdfPath: string): string[] {
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract using MiniCPM-V via Ollama
|
||||
* Query MiniCPM-V with a prompt (thinking disabled for speed)
|
||||
*/
|
||||
async function extractWithMiniCPM(images: string[], passLabel: string): Promise<ITransaction[]> {
|
||||
const payload = {
|
||||
model: MINICPM_MODEL,
|
||||
prompt: MINICPM_EXTRACT_PROMPT,
|
||||
images,
|
||||
stream: true,
|
||||
options: {
|
||||
num_predict: 16384,
|
||||
temperature: 0.1,
|
||||
},
|
||||
};
|
||||
|
||||
async function queryVision(image: string, prompt: string): Promise<string> {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/generate`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify(payload),
|
||||
body: JSON.stringify({
|
||||
model: MODEL,
|
||||
prompt: `/no_think\n${prompt}`,
|
||||
images: [image],
|
||||
stream: false,
|
||||
options: {
|
||||
num_predict: 500,
|
||||
temperature: 0.1,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
if (!response.ok) {
|
||||
throw new Error(`Ollama API error: ${response.status}`);
|
||||
}
|
||||
|
||||
const reader = response.body?.getReader();
|
||||
if (!reader) {
|
||||
throw new Error('No response body');
|
||||
}
|
||||
const data = await response.json();
|
||||
return (data.response || '').trim();
|
||||
}
|
||||
|
||||
const decoder = new TextDecoder();
|
||||
let fullText = '';
|
||||
let lineBuffer = '';
|
||||
/**
|
||||
* Count transactions on a page
|
||||
*/
|
||||
async function countTransactions(image: string, pageNum: number): Promise<number> {
|
||||
const response = await queryVision(image,
|
||||
`Count the transaction rows in this bank statement table.
|
||||
Each transaction has a date, description, and amount (debit or credit).
|
||||
Do not count headers or totals.
|
||||
How many transaction rows are there? Answer with just the number.`
|
||||
);
|
||||
|
||||
console.log(`[${passLabel}] Extracting with MiniCPM-V...`);
|
||||
console.log(` [Page ${pageNum}] Count response: "${response}"`);
|
||||
const match = response.match(/(\d+)/);
|
||||
const count = match ? parseInt(match[1], 10) : 0;
|
||||
console.log(` [Page ${pageNum}] Parsed count: ${count}`);
|
||||
return count;
|
||||
}
|
||||
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
/**
|
||||
* Get a single transaction by index (logs immediately)
|
||||
*/
|
||||
async function getTransaction(image: string, index: number, pageNum: number): Promise<ITransaction | null> {
|
||||
const response = await queryVision(image,
|
||||
`Look at transaction row #${index} in the bank statement table (row 1 is the first transaction after the header).
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
const lines = chunk.split('\n').filter((l) => l.trim());
|
||||
Extract:
|
||||
- DATE: in YYYY-MM-DD format
|
||||
- COUNTERPARTY: the description/name
|
||||
- AMOUNT: as a number (negative for debits like "- 21,47 €" = -21.47, positive for credits)
|
||||
|
||||
for (const line of lines) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
if (json.response) {
|
||||
fullText += json.response;
|
||||
lineBuffer += json.response;
|
||||
Format your answer as: DATE|COUNTERPARTY|AMOUNT
|
||||
Example: 2024-01-15|Amazon|-25.99`
|
||||
);
|
||||
|
||||
if (lineBuffer.includes('\n')) {
|
||||
const parts = lineBuffer.split('\n');
|
||||
for (let i = 0; i < parts.length - 1; i++) {
|
||||
console.log(parts[i]);
|
||||
}
|
||||
lineBuffer = parts[parts.length - 1];
|
||||
}
|
||||
}
|
||||
} catch {
|
||||
// Skip invalid JSON lines
|
||||
}
|
||||
// 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;
|
||||
}
|
||||
|
||||
if (lineBuffer) {
|
||||
console.log(lineBuffer);
|
||||
}
|
||||
console.log('');
|
||||
|
||||
const startIdx = fullText.indexOf('[');
|
||||
const endIdx = fullText.lastIndexOf(']') + 1;
|
||||
|
||||
if (startIdx < 0 || endIdx <= startIdx) {
|
||||
throw new Error('No JSON array found in response');
|
||||
}
|
||||
|
||||
return JSON.parse(fullText.substring(startIdx, endIdx));
|
||||
// Log raw response on parse failure
|
||||
console.log(` [P${pageNum} Tx${index.toString().padStart(2, ' ')}] PARSE FAILED: "${response.replace(/\n/g, ' ').substring(0, 60)}..."`);
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* Create a hash of transactions for comparison
|
||||
* Extract transactions from a single page using multi-query approach
|
||||
*/
|
||||
function hashTransactions(transactions: ITransaction[]): string {
|
||||
return transactions
|
||||
.map((t) => `${t.date}|${t.amount.toFixed(2)}`)
|
||||
.sort()
|
||||
.join(';');
|
||||
}
|
||||
async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
|
||||
// Step 1: Count transactions
|
||||
const count = await countTransactions(image, pageNum);
|
||||
|
||||
/**
|
||||
* Extract with consensus voting using MiniCPM-V only
|
||||
*/
|
||||
async function extractWithConsensus(
|
||||
images: string[],
|
||||
maxPasses: number = 5
|
||||
): Promise<ITransaction[]> {
|
||||
const results: Array<{ transactions: ITransaction[]; hash: string }> = [];
|
||||
const hashCounts: Map<string, number> = new Map();
|
||||
if (count === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const addResult = (transactions: ITransaction[], passLabel: string): number => {
|
||||
const hash = hashTransactions(transactions);
|
||||
results.push({ transactions, hash });
|
||||
hashCounts.set(hash, (hashCounts.get(hash) || 0) + 1);
|
||||
console.log(
|
||||
`[${passLabel}] Got ${transactions.length} transactions (hash: ${hash.substring(0, 20)}...)`
|
||||
// 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))
|
||||
);
|
||||
return hashCounts.get(hash)!;
|
||||
};
|
||||
|
||||
console.log('[Setup] Using MiniCPM-V only');
|
||||
|
||||
for (let pass = 1; pass <= maxPasses; pass++) {
|
||||
try {
|
||||
const transactions = await extractWithMiniCPM(images, `Pass ${pass} MiniCPM-V`);
|
||||
const count = addResult(transactions, `Pass ${pass} MiniCPM-V`);
|
||||
|
||||
if (count >= 2) {
|
||||
console.log(`[Consensus] Reached after ${pass} passes`);
|
||||
return transactions;
|
||||
for (const tx of results) {
|
||||
if (tx) {
|
||||
transactions.push(tx);
|
||||
}
|
||||
|
||||
console.log(`[Pass ${pass}] No consensus yet, trying again...`);
|
||||
} catch (err) {
|
||||
console.log(`[Pass ${pass}] Error: ${err}`);
|
||||
}
|
||||
}
|
||||
|
||||
// No consensus reached - return the most common result
|
||||
let bestHash = '';
|
||||
let bestCount = 0;
|
||||
for (const [hash, count] of hashCounts) {
|
||||
if (count > bestCount) {
|
||||
bestCount = count;
|
||||
bestHash = hash;
|
||||
}
|
||||
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 ${images.length} page(s) with MiniCPM-V (multi-query, deep think)`);
|
||||
|
||||
const allTransactions: ITransaction[] = [];
|
||||
|
||||
for (let i = 0; i < images.length; i++) {
|
||||
const pageTransactions = await extractTransactionsFromPage(images[i], i + 1);
|
||||
allTransactions.push(...pageTransactions);
|
||||
}
|
||||
|
||||
if (!bestHash) {
|
||||
throw new Error('No valid results obtained');
|
||||
}
|
||||
|
||||
const best = results.find((r) => r.hash === bestHash)!;
|
||||
console.log(`[No consensus] Using most common result (${bestCount}/${maxPasses} passes)`);
|
||||
return best.transactions;
|
||||
console.log(` [Vision] Total: ${allTransactions.length} transactions`);
|
||||
return allTransactions;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -273,62 +256,69 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
|
||||
}
|
||||
}
|
||||
|
||||
return testCases;
|
||||
return testCases.sort((a, b) => a.name.localeCompare(b.name));
|
||||
}
|
||||
|
||||
// Tests
|
||||
|
||||
tap.test('setup: ensure Docker containers are running', async () => {
|
||||
console.log('\n[Setup] Checking Docker containers...\n');
|
||||
|
||||
// Ensure MiniCPM is running
|
||||
const minicpmOk = await ensureMiniCpm();
|
||||
expect(minicpmOk).toBeTrue();
|
||||
|
||||
console.log('\n[Setup] All containers ready!\n');
|
||||
});
|
||||
|
||||
tap.test('should have MiniCPM-V 4.5 model loaded', async () => {
|
||||
tap.test('should have MiniCPM-V model loaded', async () => {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/tags`);
|
||||
const data = await response.json();
|
||||
const modelNames = data.models.map((m: { name: string }) => m.name);
|
||||
expect(modelNames.some((name: string) => name.includes('minicpm-v4.5'))).toBeTrue();
|
||||
expect(modelNames.some((name: string) => name.includes('minicpm'))).toBeTrue();
|
||||
});
|
||||
|
||||
// Dynamic test for each PDF/JSON pair
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} bank statement test cases (MiniCPM-V only)\n`);
|
||||
console.log(`\nFound ${testCases.length} bank statement test cases (MiniCPM-V)\n`);
|
||||
|
||||
let passedCount = 0;
|
||||
let failedCount = 0;
|
||||
|
||||
for (const testCase of testCases) {
|
||||
tap.test(`should extract transactions from ${testCase.name}`, async () => {
|
||||
// Load expected transactions
|
||||
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`);
|
||||
|
||||
// Convert PDF to images
|
||||
console.log('Converting PDF to images...');
|
||||
const images = convertPdfToImages(testCase.pdfPath);
|
||||
console.log(`Converted: ${images.length} pages\n`);
|
||||
console.log(` Pages: ${images.length}`);
|
||||
|
||||
// Extract with consensus (MiniCPM-V only)
|
||||
const extracted = await extractWithConsensus(images);
|
||||
console.log(`\nFinal: ${extracted.length} transactions`);
|
||||
const extracted = await extractTransactions(images);
|
||||
console.log(` Extracted: ${extracted.length} transactions`);
|
||||
|
||||
// Compare results
|
||||
const result = compareTransactions(extracted, expected);
|
||||
console.log(`Accuracy: ${result.matches}/${result.total}`);
|
||||
const accuracy = result.total > 0 ? result.matches / result.total : 0;
|
||||
|
||||
if (result.errors.length > 0) {
|
||||
console.log('Errors:');
|
||||
result.errors.forEach((e) => console.log(` - ${e}`));
|
||||
if (accuracy >= 0.95 && extracted.length === expected.length) {
|
||||
passedCount++;
|
||||
console.log(` Result: PASS (${result.matches}/${result.total})`);
|
||||
} else {
|
||||
failedCount++;
|
||||
console.log(` Result: FAIL (${result.matches}/${result.total})`);
|
||||
result.errors.slice(0, 5).forEach((e) => console.log(` - ${e}`));
|
||||
}
|
||||
|
||||
// Assert high accuracy
|
||||
const accuracy = result.matches / result.total;
|
||||
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 (MiniCPM-V)`);
|
||||
console.log(`======================================================`);
|
||||
console.log(` Method: Multi-query (no_think)`);
|
||||
console.log(` Passed: ${passedCount}/${total}`);
|
||||
console.log(` Failed: ${failedCount}/${total}`);
|
||||
console.log(`======================================================\n`);
|
||||
});
|
||||
|
||||
export default tap.start();
|
||||
|
||||
Reference in New Issue
Block a user