531 lines
17 KiB
TypeScript
531 lines
17 KiB
TypeScript
/**
|
||
* Bank statement extraction using Nanonets-OCR-s + Qwen3 (two-stage pipeline)
|
||
*
|
||
* Stage 1: Nanonets-OCR-s converts document pages to markdown (its strength)
|
||
* Stage 2: Qwen3 extracts structured JSON from the combined markdown
|
||
*
|
||
* This leverages each model's strengths:
|
||
* - Nanonets: Document OCR with semantic tags
|
||
* - Qwen3: Text understanding and JSON extraction
|
||
*/
|
||
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 { ensureNanonetsOcr, ensureMiniCpm } from './helpers/docker.js';
|
||
|
||
const NANONETS_URL = 'http://localhost:8000/v1';
|
||
const NANONETS_MODEL = 'nanonets/Nanonets-OCR-s';
|
||
|
||
const OLLAMA_URL = 'http://localhost:11434';
|
||
const QWEN_MODEL = 'qwen3:8b';
|
||
|
||
interface ITransaction {
|
||
date: string;
|
||
counterparty: string;
|
||
amount: number;
|
||
}
|
||
|
||
// Nanonets-specific prompt for document OCR to markdown
|
||
const NANONETS_OCR_PROMPT = `Extract the text from the above document as if you were reading it naturally.
|
||
Return the tables in html format.
|
||
Return the equations in LaTeX representation.
|
||
If there is an image in the document and image caption is not present, add a small description inside <img></img> tag.
|
||
Watermarks should be wrapped in brackets. Ex: <watermark>OFFICIAL COPY</watermark>.
|
||
Page numbers should be wrapped in brackets. Ex: <page_number>14</page_number>.`;
|
||
|
||
// JSON extraction prompt for Qwen3
|
||
const JSON_EXTRACTION_PROMPT = `You are a financial data extractor. Below is a bank statement converted to text/markdown. Extract ALL transactions from it 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}
|
||
]
|
||
|
||
BANK STATEMENT TEXT:
|
||
`;
|
||
|
||
/**
|
||
* 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 {
|
||
// Use 150 DPI to keep images within model's context length
|
||
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 });
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Stage 1: Convert a single page to markdown using Nanonets-OCR-s
|
||
*/
|
||
async function convertPageToMarkdown(image: string, pageNum: number): Promise<string> {
|
||
console.log(` [Nanonets] Converting page ${pageNum} to markdown...`);
|
||
const startTime = Date.now();
|
||
|
||
const response = await fetch(`${NANONETS_URL}/chat/completions`, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
'Authorization': 'Bearer dummy',
|
||
},
|
||
body: JSON.stringify({
|
||
model: NANONETS_MODEL,
|
||
messages: [{
|
||
role: 'user',
|
||
content: [
|
||
{ type: 'image_url', image_url: { url: `data:image/png;base64,${image}` }},
|
||
{ type: 'text', text: NANONETS_OCR_PROMPT },
|
||
],
|
||
}],
|
||
max_tokens: 4096,
|
||
temperature: 0.0,
|
||
}),
|
||
});
|
||
|
||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||
|
||
if (!response.ok) {
|
||
const errorText = await response.text();
|
||
console.log(` [Nanonets] ERROR page ${pageNum}: ${response.status} - ${errorText}`);
|
||
throw new Error(`Nanonets API error: ${response.status}`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const content = (data.choices?.[0]?.message?.content || '').trim();
|
||
console.log(` [Nanonets] Page ${pageNum} converted (${elapsed}s, ${content.length} chars)`);
|
||
return content;
|
||
}
|
||
|
||
/**
|
||
* Stage 1: Convert all pages to markdown using Nanonets-OCR-s
|
||
*/
|
||
async function convertDocumentToMarkdown(images: string[]): Promise<string> {
|
||
console.log(` [Stage 1] Converting ${images.length} page(s) to markdown with Nanonets-OCR-s...`);
|
||
|
||
const markdownPages: string[] = [];
|
||
|
||
for (let i = 0; i < images.length; i++) {
|
||
const markdown = await convertPageToMarkdown(images[i], i + 1);
|
||
markdownPages.push(`--- PAGE ${i + 1} ---\n${markdown}`);
|
||
}
|
||
|
||
const fullMarkdown = markdownPages.join('\n\n');
|
||
console.log(` [Stage 1] Complete: ${fullMarkdown.length} chars total`);
|
||
return fullMarkdown;
|
||
}
|
||
|
||
/**
|
||
* Ensure Qwen3 model is available
|
||
*/
|
||
async function ensureQwen3(): Promise<boolean> {
|
||
try {
|
||
const response = await fetch(`${OLLAMA_URL}/api/tags`);
|
||
if (response.ok) {
|
||
const data = await response.json();
|
||
const models = data.models || [];
|
||
if (models.some((m: { name: string }) => m.name === QWEN_MODEL)) {
|
||
console.log(` [Ollama] Model available: ${QWEN_MODEL}`);
|
||
return true;
|
||
}
|
||
}
|
||
} catch {
|
||
return false;
|
||
}
|
||
|
||
console.log(` [Ollama] Pulling ${QWEN_MODEL}...`);
|
||
const pullResponse = await fetch(`${OLLAMA_URL}/api/pull`, {
|
||
method: 'POST',
|
||
headers: { 'Content-Type': 'application/json' },
|
||
body: JSON.stringify({ name: QWEN_MODEL, stream: false }),
|
||
});
|
||
|
||
return pullResponse.ok;
|
||
}
|
||
|
||
/**
|
||
* Stage 2: Extract transactions from markdown using Qwen3
|
||
*/
|
||
async function extractTransactionsFromMarkdown(markdown: string, queryId: string): Promise<ITransaction[]> {
|
||
console.log(` [${queryId}] Sending markdown to ${QWEN_MODEL}...`);
|
||
const startTime = Date.now();
|
||
|
||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||
method: 'POST',
|
||
headers: { 'Content-Type': 'application/json' },
|
||
body: JSON.stringify({
|
||
model: QWEN_MODEL,
|
||
messages: [{
|
||
role: 'user',
|
||
content: JSON_EXTRACTION_PROMPT + markdown,
|
||
}],
|
||
stream: false,
|
||
options: {
|
||
num_predict: 8000,
|
||
temperature: 0.1,
|
||
},
|
||
}),
|
||
});
|
||
|
||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||
|
||
if (!response.ok) {
|
||
console.log(` [${queryId}] ERROR: ${response.status} (${elapsed}s)`);
|
||
throw new Error(`Ollama API error: ${response.status}`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const content = (data.message?.content || '').trim();
|
||
console.log(` [${queryId}] Response received (${elapsed}s, ${content.length} chars)`);
|
||
|
||
return parseJsonResponse(content, queryId);
|
||
}
|
||
|
||
/**
|
||
* Sanitize JSON string - fix common issues from LLM output
|
||
*/
|
||
function sanitizeJson(jsonStr: string): string {
|
||
let s = jsonStr;
|
||
|
||
// Fix +number (e.g., +93.80 -> 93.80) - JSON doesn't allow + prefix
|
||
s = s.replace(/"amount"\s*:\s*\+/g, '"amount": ');
|
||
s = s.replace(/:\s*\+(\d)/g, ': $1');
|
||
|
||
// Fix European number format with thousands separator
|
||
s = s.replace(/"amount"\s*:\s*(-?)(\d{1,3})\.(\d{3})\.(\d{2})\b/g, '"amount": $1$2$3.$4');
|
||
|
||
// Fix trailing commas before ] or }
|
||
s = s.replace(/,\s*([}\]])/g, '$1');
|
||
|
||
// Fix unescaped newlines/tabs inside strings
|
||
s = s.replace(/"([^"\\]*)\n([^"]*)"/g, '"$1 $2"');
|
||
s = s.replace(/"([^"\\]*)\t([^"]*)"/g, '"$1 $2"');
|
||
|
||
// Remove control characters
|
||
s = s.replace(/[\x00-\x08\x0B\x0C\x0E-\x1F]/g, ' ');
|
||
|
||
return s;
|
||
}
|
||
|
||
/**
|
||
* Parse amount from various formats
|
||
*/
|
||
function parseAmount(value: unknown): number {
|
||
if (typeof value === 'number') return value;
|
||
if (typeof value !== 'string') return 0;
|
||
|
||
let s = value.replace(/[€$£\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;
|
||
}
|
||
|
||
/**
|
||
* Parse JSON response into transactions
|
||
*/
|
||
function parseJsonResponse(response: string, queryId: string): ITransaction[] {
|
||
console.log(` [${queryId}] Parsing response...`);
|
||
|
||
// Remove thinking tags if present (Qwen3 may include <think>...</think>)
|
||
let cleanResponse = response.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
|
||
|
||
// Try to find JSON in markdown code block
|
||
const codeBlockMatch = cleanResponse.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||
let jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : cleanResponse;
|
||
|
||
// Sanitize JSON
|
||
jsonStr = sanitizeJson(jsonStr);
|
||
|
||
try {
|
||
const parsed = JSON.parse(jsonStr);
|
||
if (Array.isArray(parsed)) {
|
||
const txs = parsed.map(tx => ({
|
||
date: String(tx.date || ''),
|
||
counterparty: String(tx.counterparty || tx.description || ''),
|
||
amount: parseAmount(tx.amount),
|
||
}));
|
||
console.log(` [${queryId}] Parsed ${txs.length} transactions`);
|
||
return txs;
|
||
}
|
||
} catch (e) {
|
||
console.log(` [${queryId}] Direct parse failed: ${(e as Error).message}`);
|
||
|
||
// Try to find JSON array pattern
|
||
const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
|
||
if (arrayMatch) {
|
||
try {
|
||
const parsed = JSON.parse(sanitizeJson(arrayMatch[0]));
|
||
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 ${txs.length} transactions (array match)`);
|
||
return txs;
|
||
}
|
||
} catch (e2) {
|
||
console.log(` [${queryId}] Array parse failed: ${(e2 as Error).message}`);
|
||
}
|
||
}
|
||
}
|
||
|
||
console.log(` [${queryId}] PARSE FAILED - returning empty array`);
|
||
return [];
|
||
}
|
||
|
||
/**
|
||
* Compare two transaction arrays for consensus
|
||
*/
|
||
function transactionArraysMatch(a: ITransaction[], b: ITransaction[]): boolean {
|
||
if (a.length !== b.length) return false;
|
||
|
||
for (let i = 0; i < a.length; i++) {
|
||
const dateMatch = a[i].date === b[i].date;
|
||
const amountMatch = Math.abs(a[i].amount - b[i].amount) < 0.01;
|
||
if (!dateMatch || !amountMatch) return false;
|
||
}
|
||
|
||
return true;
|
||
}
|
||
|
||
/**
|
||
* Stage 2: Extract transactions using Qwen3 with consensus
|
||
*/
|
||
async function extractWithConsensus(markdown: string): Promise<ITransaction[]> {
|
||
const MAX_ATTEMPTS = 3;
|
||
console.log(` [Stage 2] Extracting transactions with ${QWEN_MODEL} (consensus)...`);
|
||
|
||
for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt++) {
|
||
console.log(`\n [Stage 2] --- Attempt ${attempt}/${MAX_ATTEMPTS} ---`);
|
||
|
||
// Extract twice in parallel
|
||
const [txs1, txs2] = await Promise.all([
|
||
extractTransactionsFromMarkdown(markdown, `A${attempt}Q1`),
|
||
extractTransactionsFromMarkdown(markdown, `A${attempt}Q2`),
|
||
]);
|
||
|
||
console.log(` [Stage 2] Results: Q1=${txs1.length} txs, Q2=${txs2.length} txs`);
|
||
|
||
if (txs1.length > 0 && transactionArraysMatch(txs1, txs2)) {
|
||
console.log(` [Stage 2] CONSENSUS REACHED: ${txs1.length} transactions`);
|
||
return txs1;
|
||
}
|
||
|
||
console.log(` [Stage 2] NO CONSENSUS`);
|
||
}
|
||
|
||
// Fallback: use last response
|
||
console.log(`\n [Stage 2] === FALLBACK ===`);
|
||
const fallback = await extractTransactionsFromMarkdown(markdown, 'FALLBACK');
|
||
console.log(` [Stage 2] ~ FALLBACK RESULT: ${fallback.length} transactions`);
|
||
return fallback;
|
||
}
|
||
|
||
/**
|
||
* Full pipeline: PDF -> Images -> Markdown -> JSON
|
||
*/
|
||
async function extractTransactions(images: string[]): Promise<ITransaction[]> {
|
||
// Stage 1: Convert to markdown
|
||
const markdown = await convertDocumentToMarkdown(images);
|
||
|
||
// Stage 2: Extract transactions with consensus
|
||
const transactions = await extractWithConsensus(markdown);
|
||
|
||
// Log all transactions
|
||
console.log(`\n [Result] 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;
|
||
}
|
||
|
||
/**
|
||
* 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} ${exp.counterparty}`);
|
||
continue;
|
||
}
|
||
|
||
const dateMatch = ext.date === exp.date;
|
||
const amountMatch = Math.abs(ext.amount - exp.amount) < 0.01;
|
||
|
||
if (dateMatch && amountMatch) {
|
||
matches++;
|
||
if (ext.counterparty !== exp.counterparty) {
|
||
variations.push(`[${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 containers are running', async () => {
|
||
console.log('\n[Setup] Checking Docker containers...\n');
|
||
|
||
// Nanonets for OCR
|
||
const nanonetsOk = await ensureNanonetsOcr();
|
||
expect(nanonetsOk).toBeTrue();
|
||
|
||
// Ollama for Qwen3
|
||
const ollamaOk = await ensureMiniCpm();
|
||
expect(ollamaOk).toBeTrue();
|
||
|
||
// Qwen3 model
|
||
const qwenOk = await ensureQwen3();
|
||
expect(qwenOk).toBeTrue();
|
||
|
||
console.log('\n[Setup] All containers ready!\n');
|
||
});
|
||
|
||
tap.test('should have models available', async () => {
|
||
// Check Nanonets
|
||
const nanonetsResp = await fetch(`${NANONETS_URL}/models`);
|
||
expect(nanonetsResp.ok).toBeTrue();
|
||
|
||
// Check Qwen3
|
||
const ollamaResp = await fetch(`${OLLAMA_URL}/api/tags`);
|
||
expect(ollamaResp.ok).toBeTrue();
|
||
const data = await ollamaResp.json();
|
||
const modelNames = data.models.map((m: { name: string }) => m.name);
|
||
expect(modelNames.some((name: string) => name.includes('qwen3'))).toBeTrue();
|
||
});
|
||
|
||
const testCases = findTestCases();
|
||
console.log(`\nFound ${testCases.length} bank statement test cases (Nanonets + Qwen3)\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(` - ${e}`));
|
||
}
|
||
|
||
if (result.variations.length > 0) {
|
||
console.log(` Counterparty variations (${result.variations.length}):`);
|
||
result.variations.slice(0, 5).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 (Nanonets + Qwen3 Pipeline)`);
|
||
console.log(`======================================================`);
|
||
console.log(` Stage 1: Nanonets-OCR-s (document -> markdown)`);
|
||
console.log(` Stage 2: Qwen3 8B (markdown -> JSON)`);
|
||
console.log(` Passed: ${passedCount}/${total}`);
|
||
console.log(` Failed: ${failedCount}/${total}`);
|
||
console.log(`======================================================\n`);
|
||
});
|
||
|
||
export default tap.start();
|