Juergen Kunz
d384c1d79b
- Update test.invoices.nanonets.ts to use DualAgentOrchestrator for JSON extraction - Enable streaming token callback for real-time progress visibility - Add markdown caching to avoid re-running Nanonets OCR for cached files - Update test.bankstatements.minicpm.ts and test.invoices.minicpm.ts with streaming - Update dependencies to @push.rocks/smartai@0.11.1 and @push.rocks/smartagent@1.2.8
489 lines
16 KiB
TypeScript
489 lines
16 KiB
TypeScript
/**
|
||
* Bank statement extraction using MiniCPM-V (visual extraction)
|
||
*
|
||
* JSON per-page approach with streaming output:
|
||
* 1. Ask for structured JSON of all transactions per page
|
||
* 2. Single pass extraction (no consensus)
|
||
*/
|
||
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 MODEL = 'openbmb/minicpm-v4.5:q8_0';
|
||
|
||
interface ITransaction {
|
||
date: string;
|
||
counterparty: string;
|
||
amount: number;
|
||
}
|
||
|
||
const JSON_PROMPT = `Extract ALL transactions from this bank statement page as a JSON array.
|
||
|
||
IMPORTANT RULES:
|
||
1. Each transaction has: date, description/counterparty, and an amount
|
||
2. Amount is NEGATIVE for money going OUT (debits, payments, withdrawals)
|
||
3. Amount is POSITIVE for money coming IN (credits, deposits, refunds)
|
||
4. Date format: YYYY-MM-DD
|
||
5. Do NOT include: opening balance, closing balance, subtotals, headers, or summary rows
|
||
6. Only include actual transactions with a specific date and amount
|
||
|
||
Return ONLY this JSON format, no explanation:
|
||
[
|
||
{"date": "2021-06-01", "counterparty": "COMPANY NAME", "amount": -25.99},
|
||
{"date": "2021-06-02", "counterparty": "DEPOSIT FROM", "amount": 100.00}
|
||
]`;
|
||
|
||
/**
|
||
* 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 {
|
||
execSync(
|
||
`convert -density 300 -quality 100 "${pdfPath}" -background white -alpha remove "${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 });
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Query for JSON extraction with streaming output
|
||
*/
|
||
async function queryJson(image: string, queryId: string): Promise<string> {
|
||
const startTime = Date.now();
|
||
process.stdout.write(` [${queryId}] `);
|
||
|
||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||
method: 'POST',
|
||
headers: { 'Content-Type': 'application/json' },
|
||
body: JSON.stringify({
|
||
model: MODEL,
|
||
messages: [{
|
||
role: 'user',
|
||
content: JSON_PROMPT,
|
||
images: [image],
|
||
}],
|
||
stream: true,
|
||
options: {
|
||
num_ctx: 32768,
|
||
num_predict: 4000,
|
||
temperature: 0.1,
|
||
},
|
||
}),
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||
process.stdout.write(`ERROR: ${response.status} (${elapsed}s)\n`);
|
||
throw new Error(`Ollama API error: ${response.status}`);
|
||
}
|
||
|
||
let content = '';
|
||
const reader = response.body!.getReader();
|
||
const decoder = new TextDecoder();
|
||
|
||
try {
|
||
while (true) {
|
||
const { done, value } = await reader.read();
|
||
if (done) break;
|
||
|
||
const chunk = decoder.decode(value, { stream: true });
|
||
for (const line of chunk.split('\n').filter(l => l.trim())) {
|
||
try {
|
||
const json = JSON.parse(line);
|
||
const token = json.message?.content || '';
|
||
if (token) {
|
||
process.stdout.write(token);
|
||
content += token;
|
||
}
|
||
} catch {
|
||
// Ignore parse errors for partial chunks
|
||
}
|
||
}
|
||
}
|
||
} finally {
|
||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||
process.stdout.write(` (${elapsed}s)\n`);
|
||
}
|
||
|
||
return content.trim();
|
||
}
|
||
|
||
/**
|
||
* Sanitize JSON string - fix common issues from vision model output
|
||
*/
|
||
function sanitizeJson(jsonStr: string): string {
|
||
let s = jsonStr;
|
||
|
||
// Fix +number (e.g., +93.80 -> 93.80) - JSON doesn't allow + prefix
|
||
// Handle various whitespace patterns
|
||
s = s.replace(/"amount"\s*:\s*\+/g, '"amount": ');
|
||
s = s.replace(/:\s*\+(\d)/g, ': $1');
|
||
|
||
// Fix European number format with thousands separator (e.g., 1.000.00 -> 1000.00)
|
||
// Pattern: "amount": X.XXX.XX where X.XXX is thousands and .XX is decimal
|
||
s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3.$4');
|
||
// Also handle larger numbers like 10.000.00
|
||
s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3$4.$5');
|
||
|
||
// Fix trailing commas before ] or }
|
||
s = s.replace(/,\s*([}\]])/g, '$1');
|
||
|
||
// Fix unescaped newlines inside strings (replace with space)
|
||
s = s.replace(/"([^"\\]*)\n([^"]*)"/g, '"$1 $2"');
|
||
|
||
// Fix unescaped tabs inside strings
|
||
s = s.replace(/"([^"\\]*)\t([^"]*)"/g, '"$1 $2"');
|
||
|
||
// Fix unescaped backslashes (but not already escaped ones)
|
||
s = s.replace(/\\(?!["\\/bfnrtu])/g, '\\\\');
|
||
|
||
// Fix common issues with counterparty names containing special chars
|
||
s = s.replace(/"counterparty":\s*"([^"]*)'([^"]*)"/g, '"counterparty": "$1$2"');
|
||
|
||
// Remove control characters except newlines (which we handle above)
|
||
s = s.replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, ' ');
|
||
|
||
return s;
|
||
}
|
||
|
||
/**
|
||
* Parse JSON response into transactions
|
||
*/
|
||
function parseJsonResponse(response: string, queryId: string): ITransaction[] {
|
||
console.log(` [${queryId}] Parsing response...`);
|
||
|
||
// Try to find JSON in markdown code block
|
||
const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||
let jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
|
||
|
||
if (codeBlockMatch) {
|
||
console.log(` [${queryId}] Found JSON in code block`);
|
||
}
|
||
|
||
// Sanitize JSON (fix +number issue)
|
||
jsonStr = sanitizeJson(jsonStr);
|
||
|
||
try {
|
||
const parsed = JSON.parse(jsonStr);
|
||
if (Array.isArray(parsed)) {
|
||
const txs = parsed.map(tx => ({
|
||
date: String(tx.date || ''),
|
||
counterparty: String(tx.counterparty || tx.description || ''),
|
||
amount: parseAmount(tx.amount),
|
||
}));
|
||
console.log(` [${queryId}] Parsed ${txs.length} transactions (direct)`);
|
||
return txs;
|
||
}
|
||
console.log(` [${queryId}] Parsed JSON is not an array`);
|
||
} catch (e) {
|
||
const errMsg = (e as Error).message;
|
||
console.log(` [${queryId}] Direct parse failed: ${errMsg}`);
|
||
|
||
// Log problematic section with context
|
||
const posMatch = errMsg.match(/position (\d+)/);
|
||
if (posMatch) {
|
||
const pos = parseInt(posMatch[1]);
|
||
const start = Math.max(0, pos - 40);
|
||
const end = Math.min(jsonStr.length, pos + 40);
|
||
const context = jsonStr.substring(start, end);
|
||
const marker = ' '.repeat(pos - start) + '^';
|
||
console.log(` [${queryId}] Context around error position ${pos}:`);
|
||
console.log(` [${queryId}] ...${context}...`);
|
||
console.log(` [${queryId}] ${marker}`);
|
||
}
|
||
|
||
// Try to find JSON array pattern
|
||
const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
|
||
if (arrayMatch) {
|
||
console.log(` [${queryId}] Found array pattern, trying to parse...`);
|
||
const sanitizedArray = sanitizeJson(arrayMatch[0]);
|
||
try {
|
||
const parsed = JSON.parse(sanitizedArray);
|
||
if (Array.isArray(parsed)) {
|
||
const txs = parsed.map(tx => ({
|
||
date: String(tx.date || ''),
|
||
counterparty: String(tx.counterparty || tx.description || ''),
|
||
amount: parseAmount(tx.amount),
|
||
}));
|
||
console.log(` [${queryId}] Parsed ${txs.length} transactions (array match)`);
|
||
return txs;
|
||
}
|
||
} catch (e2) {
|
||
const errMsg2 = (e2 as Error).message;
|
||
console.log(` [${queryId}] Array parse failed: ${errMsg2}`);
|
||
const posMatch2 = errMsg2.match(/position (\d+)/);
|
||
if (posMatch2) {
|
||
const pos2 = parseInt(posMatch2[1]);
|
||
console.log(` [${queryId}] Context around error: ...${sanitizedArray.substring(Math.max(0, pos2 - 30), pos2 + 30)}...`);
|
||
}
|
||
|
||
// Try to extract individual objects from the malformed array
|
||
console.log(` [${queryId}] Attempting object-by-object extraction...`);
|
||
const extracted = extractTransactionsFromMalformedJson(sanitizedArray, queryId);
|
||
if (extracted.length > 0) {
|
||
console.log(` [${queryId}] Recovered ${extracted.length} transactions via object extraction`);
|
||
return extracted;
|
||
}
|
||
}
|
||
} else {
|
||
console.log(` [${queryId}] No array pattern found in response`);
|
||
console.log(` [${queryId}] Raw response preview: ${response.substring(0, 200)}...`);
|
||
}
|
||
}
|
||
|
||
console.log(` [${queryId}] PARSE FAILED - returning empty array`);
|
||
return [];
|
||
}
|
||
|
||
/**
|
||
* Extract transactions from malformed JSON by parsing objects individually
|
||
*/
|
||
function extractTransactionsFromMalformedJson(jsonStr: string, queryId: string): ITransaction[] {
|
||
const transactions: ITransaction[] = [];
|
||
|
||
// Match individual transaction objects
|
||
const objectPattern = /\{\s*"date"\s*:\s*"([^"]+)"\s*,\s*"counterparty"\s*:\s*"([^"]+)"\s*,\s*"amount"\s*:\s*([+-]?\d+\.?\d*)\s*\}/g;
|
||
let match;
|
||
|
||
while ((match = objectPattern.exec(jsonStr)) !== null) {
|
||
transactions.push({
|
||
date: match[1],
|
||
counterparty: match[2],
|
||
amount: parseFloat(match[3]),
|
||
});
|
||
}
|
||
|
||
// Also try with different field orders (amount before counterparty, etc.)
|
||
if (transactions.length === 0) {
|
||
const altPattern = /\{\s*"date"\s*:\s*"([^"]+)"[^}]*"amount"\s*:\s*([+-]?\d+\.?\d*)[^}]*\}/g;
|
||
while ((match = altPattern.exec(jsonStr)) !== null) {
|
||
// Try to extract counterparty from the match
|
||
const counterpartyMatch = match[0].match(/"counterparty"\s*:\s*"([^"]+)"/);
|
||
const descMatch = match[0].match(/"description"\s*:\s*"([^"]+)"/);
|
||
transactions.push({
|
||
date: match[1],
|
||
counterparty: counterpartyMatch?.[1] || descMatch?.[1] || 'UNKNOWN',
|
||
amount: parseFloat(match[2]),
|
||
});
|
||
}
|
||
}
|
||
|
||
return transactions;
|
||
}
|
||
|
||
/**
|
||
* Parse amount from various formats
|
||
*/
|
||
function parseAmount(value: unknown): number {
|
||
if (typeof value === 'number') return value;
|
||
if (typeof value !== 'string') return 0;
|
||
|
||
let s = value.replace(/[€$£\s]/g, '').replace('−', '-').replace('–', '-');
|
||
// European format: comma is decimal
|
||
if (s.includes(',') && s.indexOf(',') > s.lastIndexOf('.')) {
|
||
s = s.replace(/\./g, '').replace(',', '.');
|
||
} else {
|
||
s = s.replace(/,/g, '');
|
||
}
|
||
return parseFloat(s) || 0;
|
||
}
|
||
|
||
/**
|
||
* Extract transactions from a single page (single pass)
|
||
*/
|
||
async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
|
||
console.log(`\n ======== Page ${pageNum} ========`);
|
||
|
||
const queryId = `P${pageNum}`;
|
||
const response = await queryJson(image, queryId);
|
||
const transactions = parseJsonResponse(response, queryId);
|
||
|
||
console.log(` [Page ${pageNum}] Extracted ${transactions.length} transactions:`);
|
||
for (let i = 0; i < transactions.length; i++) {
|
||
const tx = transactions[i];
|
||
console.log(` ${(i + 1).toString().padStart(2)}. ${tx.date} | ${tx.counterparty.substring(0, 30).padEnd(30)} | ${tx.amount >= 0 ? '+' : ''}${tx.amount.toFixed(2)}`);
|
||
}
|
||
|
||
return transactions;
|
||
}
|
||
|
||
/**
|
||
* Extract all transactions from bank statement
|
||
*/
|
||
async function extractTransactions(images: string[]): Promise<ITransaction[]> {
|
||
console.log(` [Vision] Processing ${images.length} page(s) with ${MODEL} (single pass)`);
|
||
|
||
const allTransactions: ITransaction[] = [];
|
||
|
||
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 extracted transactions against expected
|
||
*/
|
||
function compareTransactions(
|
||
extracted: ITransaction[],
|
||
expected: ITransaction[]
|
||
): { matches: number; total: number; errors: string[]; variations: string[] } {
|
||
const errors: string[] = [];
|
||
const variations: 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} ${exp.counterparty}`);
|
||
continue;
|
||
}
|
||
|
||
const dateMatch = ext.date === exp.date;
|
||
const amountMatch = Math.abs(ext.amount - exp.amount) < 0.01;
|
||
|
||
if (dateMatch && amountMatch) {
|
||
matches++;
|
||
// Track counterparty variations (date and amount match but name differs)
|
||
if (ext.counterparty !== exp.counterparty) {
|
||
variations.push(
|
||
`[${i}] "${exp.counterparty}" → "${ext.counterparty}"`
|
||
);
|
||
}
|
||
} 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: ${extracted.length - expected.length}`);
|
||
}
|
||
|
||
return { matches, total: expected.length, errors, variations };
|
||
}
|
||
|
||
/**
|
||
* Find all test cases (PDF + JSON pairs) in .nogit/
|
||
*/
|
||
function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
|
||
const testDir = path.join(process.cwd(), '.nogit');
|
||
if (!fs.existsSync(testDir)) {
|
||
return [];
|
||
}
|
||
|
||
const files = fs.readdirSync(testDir);
|
||
const pdfFiles = files.filter((f: string) => f.endsWith('.pdf'));
|
||
const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
|
||
|
||
for (const pdf of pdfFiles) {
|
||
const baseName = pdf.replace('.pdf', '');
|
||
const jsonFile = `${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));
|
||
}
|
||
|
||
// Tests
|
||
|
||
tap.test('setup: ensure Docker containers are running', async () => {
|
||
console.log('\n[Setup] Checking Docker containers...\n');
|
||
const minicpmOk = await ensureMiniCpm();
|
||
expect(minicpmOk).toBeTrue();
|
||
console.log('\n[Setup] All containers ready!\n');
|
||
});
|
||
|
||
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'))).toBeTrue();
|
||
});
|
||
|
||
const testCases = findTestCases();
|
||
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: ${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: ${extracted.length} transactions`);
|
||
|
||
const result = compareTransactions(extracted, expected);
|
||
const perfectMatch = result.matches === result.total && extracted.length === expected.length;
|
||
|
||
if (perfectMatch) {
|
||
passedCount++;
|
||
console.log(` Result: PASS (${result.matches}/${result.total})`);
|
||
} else {
|
||
failedCount++;
|
||
console.log(` Result: FAIL (${result.matches}/${result.total})`);
|
||
result.errors.slice(0, 10).forEach((e) => console.log(` - ${e}`));
|
||
}
|
||
|
||
// Log counterparty variations (names that differ but date/amount matched)
|
||
if (result.variations.length > 0) {
|
||
console.log(` Counterparty variations (${result.variations.length}):`);
|
||
result.variations.forEach((v) => console.log(` ${v}`));
|
||
}
|
||
|
||
expect(result.matches).toEqual(result.total);
|
||
expect(extracted.length).toEqual(expected.length);
|
||
});
|
||
}
|
||
|
||
tap.test('summary', async () => {
|
||
const total = testCases.length;
|
||
console.log(`\n======================================================`);
|
||
console.log(` Bank Statement Summary (${MODEL})`);
|
||
console.log(`======================================================`);
|
||
console.log(` Method: JSON per-page (single pass)`);
|
||
console.log(` Passed: ${passedCount}/${total}`);
|
||
console.log(` Failed: ${failedCount}/${total}`);
|
||
console.log(`======================================================\n`);
|
||
});
|
||
|
||
export default tap.start();
|