Compare commits
11 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| d6c97a9625 | |||
| 76b21f1f7b | |||
| 4c368dfef9 | |||
| e76768da55 | |||
| 63d72a52c9 | |||
| 386122c8c7 | |||
| 7c8f10497e | |||
| 9f9ec0a671 | |||
| 3780105c6f | |||
| d237ad19f4 | |||
| 7652a2df52 |
26
Dockerfile_qwen3vl
Normal file
26
Dockerfile_qwen3vl
Normal file
@@ -0,0 +1,26 @@
|
||||
# Qwen3-VL-30B-A3B Vision Language Model
|
||||
# Q4_K_M quantization (~20GB model)
|
||||
#
|
||||
# Most powerful Qwen vision model:
|
||||
# - 256K context (expandable to 1M)
|
||||
# - Visual agent capabilities
|
||||
# - Code generation from images
|
||||
#
|
||||
# Build: docker build -f Dockerfile_qwen3vl -t qwen3vl .
|
||||
# Run: docker run --gpus all -p 11434:11434 -v ht-ollama-models:/root/.ollama qwen3vl
|
||||
|
||||
FROM ollama/ollama:latest
|
||||
|
||||
# Pre-pull the model during build (optional - can also pull at runtime)
|
||||
# This makes the image larger but faster to start
|
||||
# RUN ollama serve & sleep 5 && ollama pull qwen3-vl:30b-a3b && pkill ollama
|
||||
|
||||
# Expose Ollama API port
|
||||
EXPOSE 11434
|
||||
|
||||
# Health check
|
||||
HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
|
||||
CMD curl -f http://localhost:11434/api/tags || exit 1
|
||||
|
||||
# Start Ollama server
|
||||
CMD ["serve"]
|
||||
48
changelog.md
48
changelog.md
@@ -1,5 +1,53 @@
|
||||
# Changelog
|
||||
|
||||
## 2026-01-18 - 1.12.0 - feat(tests)
|
||||
switch vision tests to multi-query extraction (count then per-row/field queries) and add logging/summaries
|
||||
|
||||
- Replace streaming + consensus pipeline with multi-query approach: count rows per page, then query each transaction/field individually (batched parallel queries).
|
||||
- Introduce unified helpers (queryVision / queryField / getTransaction / countTransactions) and simplify Ollama requests (stream:false, reduced num_predict, /no_think prompts).
|
||||
- Improve parsing and normalization for amounts (European formats), invoice numbers, dates and currency extraction.
|
||||
- Adjust model checks to look for generic 'minicpm' and update test names/messages; add pass/fail counters and a summary test output.
|
||||
- Remove previous consensus voting and streaming JSON accumulation logic, and add immediate per-transaction logging and batching.
|
||||
|
||||
## 2026-01-18 - 1.11.0 - feat(vision)
|
||||
process pages separately and make Qwen3-VL vision extraction more robust; add per-page parsing, safer JSON handling, reduced token usage, and multi-query invoice extraction
|
||||
|
||||
- Bank statements: split extraction into extractTransactionsFromPage and sequentially process pages to avoid thinking-token exhaustion
|
||||
- Bank statements: reduced num_predict from 8000 to 4000, send single image per request, added per-page logging and non-throwing handling for empty or non-JSON responses
|
||||
- Bank statements: catch JSON.parse errors and return empty array instead of throwing
|
||||
- Invoices: introduced queryField to request single values and perform multiple simple queries (reduces model thinking usage)
|
||||
- Invoices: reduced num_predict for invoice queries from 4000 to 500 and parse amounts robustly (handles European formats like 1.234,56)
|
||||
- Invoices: normalize currency to uppercase 3-letter code, return safe defaults (empty strings / 0) instead of nulls, and parse net/vat/total with fallbacks
|
||||
- General: simplified Ollama API error messages to avoid including response body content in thrown errors
|
||||
|
||||
## 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
|
||||
|
||||
- Add Dockerfile_qwen3vl to provide an Ollama-based image for Qwen3-VL and expose the Ollama API on port 11434
|
||||
- Introduce test/test.invoices.qwen3vl.ts and ensureQwen3Vl() helper to pull and test qwen3-vl:8b
|
||||
- Improve PDF->PNG conversion and prompt in ministral3 tests (higher DPI, max quality, sharpen) and increase num_predict from 512 to 1024
|
||||
- Simplify extraction pipeline: remove consensus voting, log single-pass results, and simplify OCR HTML sanitization/truncation logic
|
||||
|
||||
## 2026-01-18 - 1.9.0 - feat(tests)
|
||||
add Ministral 3 vision tests and improve invoice extraction pipeline to use Ollama chat schema, sanitization, and multi-page support
|
||||
|
||||
- Add new vision-based test suites for Ministral 3: test/test.invoices.ministral3.ts and test/test.bankstatements.ministral3.ts (model ministral-3:8b).
|
||||
- Introduce ensureMinistral3() helper to start/check Ollama/MiniCPM model in test/helpers/docker.ts.
|
||||
- Switch invoice extraction to use Ollama /api/chat with a JSON schema (format) and streaming support (reads message.content).
|
||||
- Improve HTML handling: sanitizeHtml() to remove OCR artifacts, concatenate multi-page HTML with page markers, and increase truncation limits.
|
||||
- Enhance response parsing: strip Markdown code fences, robustly locate JSON object boundaries, and provide clearer JSON parse errors.
|
||||
- Add PDF->PNG conversion (ImageMagick) and direct image-based extraction flow for vision model tests.
|
||||
|
||||
## 2026-01-18 - 1.8.0 - feat(paddleocr-vl)
|
||||
add structured HTML output and table parsing for PaddleOCR-VL, update API, tests, and README
|
||||
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
{
|
||||
"name": "@host.today/ht-docker-ai",
|
||||
"version": "1.8.0",
|
||||
"version": "1.12.0",
|
||||
"type": "module",
|
||||
"private": false,
|
||||
"description": "Docker images for AI vision-language models including MiniCPM-V 4.5",
|
||||
|
||||
@@ -311,9 +311,8 @@ export async function ensureOllamaModel(modelName: string): Promise<boolean> {
|
||||
if (response.ok) {
|
||||
const data = await response.json();
|
||||
const models = data.models || [];
|
||||
const exists = models.some((m: { name: string }) =>
|
||||
m.name === modelName || m.name.startsWith(modelName.split(':')[0])
|
||||
);
|
||||
// Exact match required - don't match on prefix
|
||||
const exists = models.some((m: { name: string }) => m.name === modelName);
|
||||
|
||||
if (exists) {
|
||||
console.log(`[Ollama] Model already available: ${modelName}`);
|
||||
@@ -358,3 +357,29 @@ export async function ensureQwen25(): Promise<boolean> {
|
||||
// Then ensure the Qwen2.5 model is pulled
|
||||
return ensureOllamaModel('qwen2.5:7b');
|
||||
}
|
||||
|
||||
/**
|
||||
* Ensure Ministral 3 8B model is available (for structured JSON extraction)
|
||||
* Ministral 3 has native JSON output support and OCR-style document extraction
|
||||
*/
|
||||
export async function ensureMinistral3(): Promise<boolean> {
|
||||
// First ensure the Ollama service (MiniCPM container) is running
|
||||
const ollamaOk = await ensureMiniCpm();
|
||||
if (!ollamaOk) return false;
|
||||
|
||||
// Then ensure the Ministral 3 8B model is pulled
|
||||
return ensureOllamaModel('ministral-3:8b');
|
||||
}
|
||||
|
||||
/**
|
||||
* Ensure Qwen3-VL 8B model is available (vision-language model)
|
||||
* Q4_K_M quantization (~5GB) - fits in 15GB VRAM with room to spare
|
||||
*/
|
||||
export async function ensureQwen3Vl(): Promise<boolean> {
|
||||
// First ensure the Ollama service is running
|
||||
const ollamaOk = await ensureMiniCpm();
|
||||
if (!ollamaOk) return false;
|
||||
|
||||
// Then ensure Qwen3-VL 8B is pulled
|
||||
return ensureOllamaModel('qwen3-vl:8b');
|
||||
}
|
||||
|
||||
@@ -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();
|
||||
|
||||
348
test/test.bankstatements.ministral3.ts
Normal file
348
test/test.bankstatements.ministral3.ts
Normal file
@@ -0,0 +1,348 @@
|
||||
/**
|
||||
* Bank Statement extraction using Ministral 3 Vision (Direct)
|
||||
*
|
||||
* NO OCR pipeline needed - Ministral 3 has built-in vision encoder:
|
||||
* 1. Convert PDF to images
|
||||
* 2. Send images directly to Ministral 3 via Ollama
|
||||
* 3. Extract transactions as structured JSON
|
||||
*/
|
||||
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 { ensureMinistral3 } from './helpers/docker.js';
|
||||
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
const VISION_MODEL = 'ministral-3:8b';
|
||||
|
||||
interface ITransaction {
|
||||
date: string;
|
||||
counterparty: string;
|
||||
amount: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* 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 200 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
|
||||
{ stdio: 'pipe' }
|
||||
);
|
||||
|
||||
const files = fs.readdirSync(tempDir).filter((f) => 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 });
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract transactions from a single page image using Ministral 3 Vision
|
||||
*/
|
||||
async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
|
||||
console.log(` [Vision] Processing page ${pageNum}`);
|
||||
|
||||
// JSON schema for array of transactions
|
||||
const transactionSchema = {
|
||||
type: 'array',
|
||||
items: {
|
||||
type: 'object',
|
||||
properties: {
|
||||
date: { type: 'string', description: 'Transaction date in YYYY-MM-DD format' },
|
||||
counterparty: { type: 'string', description: 'Name of the other party' },
|
||||
amount: { type: 'number', description: 'Amount (negative for debits, positive for credits)' },
|
||||
},
|
||||
required: ['date', 'counterparty', 'amount'],
|
||||
},
|
||||
};
|
||||
|
||||
const prompt = `Extract ALL bank transactions from this bank statement page.
|
||||
|
||||
For each transaction, extract:
|
||||
- date: Transaction date in YYYY-MM-DD format
|
||||
- counterparty: The name/description of the other party (merchant, payee, etc.)
|
||||
- amount: The amount as a number (NEGATIVE for debits/expenses, POSITIVE for credits/income)
|
||||
|
||||
Return a JSON array of transactions. If no transactions visible, return empty array [].
|
||||
Example: [{"date":"2021-06-01","counterparty":"AMAZON","amount":-50.00}]`;
|
||||
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({
|
||||
model: VISION_MODEL,
|
||||
messages: [
|
||||
{
|
||||
role: 'user',
|
||||
content: prompt,
|
||||
images: [image],
|
||||
},
|
||||
],
|
||||
format: transactionSchema,
|
||||
stream: true,
|
||||
options: {
|
||||
num_predict: 4096, // Bank statements can have many transactions
|
||||
temperature: 0.0,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
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 decoder = new TextDecoder();
|
||||
let fullText = '';
|
||||
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
const lines = chunk.split('\n').filter((l) => l.trim());
|
||||
|
||||
for (const line of lines) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
if (json.message?.content) {
|
||||
fullText += json.message.content;
|
||||
}
|
||||
} catch {
|
||||
// Skip invalid JSON lines
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 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();
|
||||
|
||||
// Find array boundaries
|
||||
const startIdx = jsonStr.indexOf('[');
|
||||
const endIdx = jsonStr.lastIndexOf(']') + 1;
|
||||
|
||||
if (startIdx < 0 || endIdx <= startIdx) {
|
||||
console.log(` [Page ${pageNum}] No transactions found`);
|
||||
return [];
|
||||
}
|
||||
|
||||
try {
|
||||
const parsed = JSON.parse(jsonStr.substring(startIdx, endIdx));
|
||||
console.log(` [Page ${pageNum}] Found ${parsed.length} transactions`);
|
||||
return parsed.map((t: { date?: string; counterparty?: string; amount?: number }) => ({
|
||||
date: t.date || '',
|
||||
counterparty: t.counterparty || '',
|
||||
amount: parseFloat(String(t.amount)) || 0,
|
||||
}));
|
||||
} catch (e) {
|
||||
console.log(` [Page ${pageNum}] Parse error: ${e}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract all transactions from all pages
|
||||
*/
|
||||
async function extractAllTransactions(images: string[]): Promise<ITransaction[]> {
|
||||
const allTransactions: ITransaction[] = [];
|
||||
|
||||
for (let i = 0; i < images.length; i++) {
|
||||
const pageTransactions = await extractTransactionsFromPage(images[i], i + 1);
|
||||
allTransactions.push(...pageTransactions);
|
||||
}
|
||||
|
||||
return allTransactions;
|
||||
}
|
||||
|
||||
/**
|
||||
* Normalize date to YYYY-MM-DD
|
||||
*/
|
||||
function normalizeDate(dateStr: string): string {
|
||||
if (!dateStr) return '';
|
||||
if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
|
||||
|
||||
// Handle DD/MM/YYYY or DD.MM.YYYY
|
||||
const match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
|
||||
if (match) {
|
||||
return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
return dateStr;
|
||||
}
|
||||
|
||||
/**
|
||||
* Compare extracted transactions vs expected
|
||||
*/
|
||||
function compareTransactions(
|
||||
extracted: ITransaction[],
|
||||
expected: ITransaction[]
|
||||
): { matchRate: number; matched: number; missed: number; extra: number; errors: string[] } {
|
||||
const errors: string[] = [];
|
||||
let matched = 0;
|
||||
|
||||
// Normalize all dates
|
||||
const normalizedExtracted = extracted.map((t) => ({
|
||||
...t,
|
||||
date: normalizeDate(t.date),
|
||||
counterparty: t.counterparty.toUpperCase().trim(),
|
||||
}));
|
||||
|
||||
const normalizedExpected = expected.map((t) => ({
|
||||
...t,
|
||||
date: normalizeDate(t.date),
|
||||
counterparty: t.counterparty.toUpperCase().trim(),
|
||||
}));
|
||||
|
||||
// Try to match each expected transaction
|
||||
const matchedIndices = new Set<number>();
|
||||
|
||||
for (const exp of normalizedExpected) {
|
||||
let found = false;
|
||||
|
||||
for (let i = 0; i < normalizedExtracted.length; i++) {
|
||||
if (matchedIndices.has(i)) continue;
|
||||
|
||||
const ext = normalizedExtracted[i];
|
||||
|
||||
// Match by date + amount (counterparty names can vary)
|
||||
if (ext.date === exp.date && Math.abs(ext.amount - exp.amount) < 0.02) {
|
||||
matched++;
|
||||
matchedIndices.add(i);
|
||||
found = true;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (!found) {
|
||||
errors.push(`Missing: ${exp.date} | ${exp.counterparty} | ${exp.amount}`);
|
||||
}
|
||||
}
|
||||
|
||||
const missed = expected.length - matched;
|
||||
const extra = extracted.length - matched;
|
||||
const matchRate = expected.length > 0 ? (matched / expected.length) * 100 : 0;
|
||||
|
||||
return { matchRate, matched, missed, extra, errors };
|
||||
}
|
||||
|
||||
/**
|
||||
* Find test cases (PDF + JSON pairs in .nogit/)
|
||||
*/
|
||||
function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
|
||||
const testDir = path.join(process.cwd(), '.nogit');
|
||||
if (!fs.existsSync(testDir)) return [];
|
||||
|
||||
const files = fs.readdirSync(testDir);
|
||||
const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
|
||||
|
||||
for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
|
||||
const baseName = pdf.replace('.pdf', '');
|
||||
const jsonFile = `${baseName}.json`;
|
||||
if (files.includes(jsonFile)) {
|
||||
// Skip invoice files - only bank statements
|
||||
if (!baseName.includes('invoice')) {
|
||||
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 Ministral 3 is running', async () => {
|
||||
console.log('\n[Setup] Checking Ministral 3...\n');
|
||||
const ok = await ensureMinistral3();
|
||||
expect(ok).toBeTrue();
|
||||
console.log('\n[Setup] Ready!\n');
|
||||
});
|
||||
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} bank statement test cases (Ministral 3 Vision)\n`);
|
||||
|
||||
let totalMatched = 0;
|
||||
let totalExpected = 0;
|
||||
const times: number[] = [];
|
||||
|
||||
for (const testCase of testCases) {
|
||||
tap.test(`should extract bank statement: ${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 start = Date.now();
|
||||
const images = convertPdfToImages(testCase.pdfPath);
|
||||
console.log(` Pages: ${images.length}`);
|
||||
|
||||
const extracted = await extractAllTransactions(images);
|
||||
const elapsed = Date.now() - start;
|
||||
times.push(elapsed);
|
||||
|
||||
console.log(` Extracted: ${extracted.length} transactions`);
|
||||
|
||||
const result = compareTransactions(extracted, expected);
|
||||
totalMatched += result.matched;
|
||||
totalExpected += expected.length;
|
||||
|
||||
console.log(` Match rate: ${result.matchRate.toFixed(1)}% (${result.matched}/${expected.length})`);
|
||||
console.log(` Missed: ${result.missed}, Extra: ${result.extra}`);
|
||||
console.log(` Time: ${(elapsed / 1000).toFixed(1)}s`);
|
||||
|
||||
if (result.errors.length > 0 && result.errors.length <= 5) {
|
||||
result.errors.forEach((e) => console.log(` - ${e}`));
|
||||
} else if (result.errors.length > 5) {
|
||||
console.log(` (${result.errors.length} missing transactions)`);
|
||||
}
|
||||
|
||||
// Consider it a pass if we match at least 70% of transactions
|
||||
expect(result.matchRate).toBeGreaterThan(70);
|
||||
});
|
||||
}
|
||||
|
||||
tap.test('summary', async () => {
|
||||
const overallMatchRate = totalExpected > 0 ? (totalMatched / totalExpected) * 100 : 0;
|
||||
const totalTime = times.reduce((a, b) => a + b, 0) / 1000;
|
||||
const avgTime = times.length > 0 ? totalTime / times.length : 0;
|
||||
|
||||
console.log(`\n======================================================`);
|
||||
console.log(` Bank Statement Extraction Summary (Ministral 3)`);
|
||||
console.log(`======================================================`);
|
||||
console.log(` Method: Ministral 3 8B Vision (Direct)`);
|
||||
console.log(` Statements: ${testCases.length}`);
|
||||
console.log(` Matched: ${totalMatched}/${totalExpected} transactions`);
|
||||
console.log(` Match rate: ${overallMatchRate.toFixed(1)}%`);
|
||||
console.log(`------------------------------------------------------`);
|
||||
console.log(` Total time: ${totalTime.toFixed(1)}s`);
|
||||
console.log(` Avg per stmt: ${avgTime.toFixed(1)}s`);
|
||||
console.log(`======================================================\n`);
|
||||
});
|
||||
|
||||
export default tap.start();
|
||||
345
test/test.bankstatements.qwen3vl.ts
Normal file
345
test/test.bankstatements.qwen3vl.ts
Normal file
@@ -0,0 +1,345 @@
|
||||
/**
|
||||
* Bank statement extraction using Qwen3-VL 8B Vision (Direct)
|
||||
*
|
||||
* Multi-query approach:
|
||||
* 1. First ask how many transactions on each page
|
||||
* 2. Then query each transaction individually
|
||||
* Single pass, no consensus voting.
|
||||
*/
|
||||
import { tap, expect } from '@git.zone/tstest/tapbundle';
|
||||
import * as fs from 'fs';
|
||||
import * as path from 'path';
|
||||
import { execSync } from 'child_process';
|
||||
import * as os from 'os';
|
||||
import { ensureMiniCpm } from './helpers/docker.js';
|
||||
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
const VISION_MODEL = 'qwen3-vl:8b';
|
||||
|
||||
interface ITransaction {
|
||||
date: string;
|
||||
counterparty: string;
|
||||
amount: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert PDF to PNG images
|
||||
*/
|
||||
function convertPdfToImages(pdfPath: string): string[] {
|
||||
const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
|
||||
const outputPattern = path.join(tempDir, 'page-%d.png');
|
||||
|
||||
try {
|
||||
execSync(
|
||||
`convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
|
||||
{ stdio: 'pipe' }
|
||||
);
|
||||
|
||||
const files = fs.readdirSync(tempDir).filter((f: string) => f.endsWith('.png')).sort();
|
||||
const images: string[] = [];
|
||||
|
||||
for (const file of files) {
|
||||
const imagePath = path.join(tempDir, file);
|
||||
const imageData = fs.readFileSync(imagePath);
|
||||
images.push(imageData.toString('base64'));
|
||||
}
|
||||
|
||||
return images;
|
||||
} finally {
|
||||
fs.rmSync(tempDir, { recursive: true, force: true });
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Query Qwen3-VL with a simple prompt
|
||||
*/
|
||||
async function queryVision(image: string, prompt: string): Promise<string> {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({
|
||||
model: VISION_MODEL,
|
||||
messages: [{
|
||||
role: 'user',
|
||||
content: prompt,
|
||||
images: [image],
|
||||
}],
|
||||
stream: false,
|
||||
options: {
|
||||
num_predict: 500,
|
||||
temperature: 0.1,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
if (!response.ok) {
|
||||
throw new Error(`Ollama API error: ${response.status}`);
|
||||
}
|
||||
|
||||
const data = await response.json();
|
||||
return (data.message?.content || '').trim();
|
||||
}
|
||||
|
||||
/**
|
||||
* 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 (multi-query)`);
|
||||
|
||||
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 transactions
|
||||
*/
|
||||
function compareTransactions(
|
||||
extracted: ITransaction[],
|
||||
expected: ITransaction[]
|
||||
): { matches: number; total: number; errors: string[] } {
|
||||
const errors: 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++;
|
||||
} 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: ${extracted.length - expected.length}`);
|
||||
}
|
||||
|
||||
return { matches, total: expected.length, errors };
|
||||
}
|
||||
|
||||
/**
|
||||
* Find test cases in .nogit/
|
||||
*/
|
||||
function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
|
||||
const testDir = path.join(process.cwd(), '.nogit');
|
||||
if (!fs.existsSync(testDir)) return [];
|
||||
|
||||
const files = fs.readdirSync(testDir);
|
||||
const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
|
||||
|
||||
for (const pdf of files.filter((f: string) => f.endsWith('.pdf'))) {
|
||||
const baseName = pdf.replace('.pdf', '');
|
||||
const jsonFile = `${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));
|
||||
}
|
||||
|
||||
/**
|
||||
* Ensure Qwen3-VL 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 available: ${VISION_MODEL}`);
|
||||
return true;
|
||||
}
|
||||
}
|
||||
} catch {
|
||||
return false;
|
||||
}
|
||||
|
||||
console.log(`[Ollama] Pulling ${VISION_MODEL}...`);
|
||||
const pullResponse = await fetch(`${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...\n');
|
||||
const ollamaOk = await ensureMiniCpm();
|
||||
expect(ollamaOk).toBeTrue();
|
||||
const visionOk = await ensureQwen3Vl();
|
||||
expect(visionOk).toBeTrue();
|
||||
console.log('\n[Setup] Ready!\n');
|
||||
});
|
||||
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} bank statement test cases (Qwen3-VL)\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: ${extracted.length} transactions`);
|
||||
|
||||
const result = compareTransactions(extracted, expected);
|
||||
const accuracy = result.total > 0 ? result.matches / result.total : 0;
|
||||
|
||||
if (accuracy >= 0.95 && extracted.length === expected.length) {
|
||||
passedCount++;
|
||||
console.log(` Result: PASS (${result.matches}/${result.total})`);
|
||||
} else {
|
||||
failedCount++;
|
||||
console.log(` Result: FAIL (${result.matches}/${result.total})`);
|
||||
result.errors.slice(0, 5).forEach((e) => console.log(` - ${e}`));
|
||||
}
|
||||
|
||||
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 (Qwen3-VL Vision)`);
|
||||
console.log(`======================================================`);
|
||||
console.log(` Method: Multi-query (count then extract each)`);
|
||||
console.log(` Passed: ${passedCount}/${total}`);
|
||||
console.log(` Failed: ${failedCount}/${total}`);
|
||||
console.log(`======================================================\n`);
|
||||
});
|
||||
|
||||
export default tap.start();
|
||||
@@ -1,8 +1,8 @@
|
||||
/**
|
||||
* Invoice extraction test using MiniCPM-V only (visual extraction)
|
||||
*
|
||||
* This tests MiniCPM-V's ability to extract invoice data directly from images
|
||||
* without any OCR augmentation.
|
||||
* Multi-query approach with thinking DISABLED for speed.
|
||||
* Single pass, no consensus voting.
|
||||
*/
|
||||
import { tap, expect } from '@git.zone/tstest/tapbundle';
|
||||
import * as fs from 'fs';
|
||||
@@ -24,28 +24,6 @@ interface IInvoice {
|
||||
total_amount: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Build extraction prompt (MiniCPM-V only, no OCR augmentation)
|
||||
*/
|
||||
function buildPrompt(): string {
|
||||
return `/nothink
|
||||
You are an invoice parser. Extract the following fields from this invoice:
|
||||
|
||||
1. invoice_number: The invoice/receipt number
|
||||
2. invoice_date: Date in YYYY-MM-DD format
|
||||
3. vendor_name: Company that issued the invoice
|
||||
4. currency: EUR, USD, etc.
|
||||
5. net_amount: Amount before tax (if shown)
|
||||
6. vat_amount: Tax/VAT amount (if shown, 0 if reverse charge or no tax)
|
||||
7. total_amount: Final amount due
|
||||
|
||||
Return ONLY valid JSON in this exact format:
|
||||
{"invoice_number":"XXX","invoice_date":"YYYY-MM-DD","vendor_name":"Company Name","currency":"EUR","net_amount":100.00,"vat_amount":19.00,"total_amount":119.00}
|
||||
|
||||
If a field is not visible, use null for strings or 0 for numbers.
|
||||
No explanation, just the JSON object.`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert PDF to PNG images using ImageMagick
|
||||
*/
|
||||
@@ -75,122 +53,312 @@ function convertPdfToImages(pdfPath: string): string[] {
|
||||
}
|
||||
|
||||
/**
|
||||
* Single extraction pass with MiniCPM-V
|
||||
* Query MiniCPM-V for a single field (thinking disabled for speed)
|
||||
*/
|
||||
async function extractOnce(images: string[], passNum: number): Promise<IInvoice> {
|
||||
const payload = {
|
||||
model: MODEL,
|
||||
prompt: buildPrompt(),
|
||||
images,
|
||||
stream: true,
|
||||
options: {
|
||||
num_predict: 2048,
|
||||
temperature: 0.1,
|
||||
},
|
||||
};
|
||||
|
||||
async function queryField(images: string[], question: string): Promise<string> {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/generate`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify(payload),
|
||||
body: JSON.stringify({
|
||||
model: MODEL,
|
||||
prompt: `/no_think\n${question}`,
|
||||
images: images,
|
||||
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();
|
||||
const content = (data.response || '').trim();
|
||||
|
||||
const decoder = new TextDecoder();
|
||||
let fullText = '';
|
||||
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
const lines = chunk.split('\n').filter((l) => l.trim());
|
||||
|
||||
for (const line of lines) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
if (json.response) {
|
||||
fullText += json.response;
|
||||
}
|
||||
} catch {
|
||||
// Skip invalid JSON lines
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Extract JSON from response
|
||||
const startIdx = fullText.indexOf('{');
|
||||
const endIdx = fullText.lastIndexOf('}') + 1;
|
||||
|
||||
if (startIdx < 0 || endIdx <= startIdx) {
|
||||
throw new Error(`No JSON object found in response: ${fullText.substring(0, 200)}`);
|
||||
}
|
||||
|
||||
const jsonStr = fullText.substring(startIdx, endIdx);
|
||||
return JSON.parse(jsonStr);
|
||||
// Return full content (no thinking to filter)
|
||||
return content;
|
||||
}
|
||||
|
||||
/**
|
||||
* Create a hash of invoice for comparison (using key fields)
|
||||
* Extract invoice data using multiple queries with validation
|
||||
*/
|
||||
function hashInvoice(invoice: IInvoice): string {
|
||||
return `${invoice.invoice_number}|${invoice.invoice_date}|${invoice.total_amount.toFixed(2)}`;
|
||||
}
|
||||
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
|
||||
console.log(` [Vision] Processing ${images.length} page(s) with MiniCPM-V (multi-query + validation)`);
|
||||
|
||||
/**
|
||||
* Extract with consensus voting using MiniCPM-V only
|
||||
*/
|
||||
async function extractWithConsensus(images: string[], invoiceName: string, maxPasses: number = 5): Promise<IInvoice> {
|
||||
const results: Array<{ invoice: IInvoice; hash: string }> = [];
|
||||
const hashCounts: Map<string, number> = new Map();
|
||||
|
||||
const addResult = (invoice: IInvoice, passLabel: string): number => {
|
||||
const hash = hashInvoice(invoice);
|
||||
results.push({ invoice, hash });
|
||||
hashCounts.set(hash, (hashCounts.get(hash) || 0) + 1);
|
||||
console.log(` [${passLabel}] ${invoice.invoice_number} | ${invoice.invoice_date} | ${invoice.total_amount} ${invoice.currency}`);
|
||||
return hashCounts.get(hash)!;
|
||||
// Log each result as it comes in
|
||||
const queryAndLog = async (name: string, question: string): Promise<string> => {
|
||||
const result = await queryField(images, question);
|
||||
console.log(` [Query] ${name}: "${result}"`);
|
||||
return result;
|
||||
};
|
||||
|
||||
for (let pass = 1; pass <= maxPasses; pass++) {
|
||||
try {
|
||||
const invoice = await extractOnce(images, pass);
|
||||
const count = addResult(invoice, `Pass ${pass}`);
|
||||
// STRATEGY 1: List-then-pick for invoice number (avoids confusion with VAT/customer IDs)
|
||||
// Also ask for invoice number directly as backup
|
||||
const [allNumbers, directInvoiceNum] = await Promise.all([
|
||||
queryAndLog('All Numbers ', `List ALL document numbers visible on this invoice.
|
||||
For each number, identify what type it is.
|
||||
Format: type:number, type:number
|
||||
Example: >>>invoice:R0014359508, vat:DE123456789, customer:K001234<<<`),
|
||||
queryAndLog('Invoice # Dir ', `What is the INVOICE NUMBER (Rechnungsnummer)?
|
||||
NOT the VAT number (starts with DE/IE), NOT customer ID.
|
||||
Look for "Invoice No.", "Rechnungsnr.", "Invoice #", or "Facture".
|
||||
For Adobe: starts with IEE or R followed by digits.
|
||||
Return ONLY the number: >>>IEE2022006460244<<<`),
|
||||
]);
|
||||
|
||||
if (count >= 2) {
|
||||
console.log(` [Consensus] Reached after ${pass} passes`);
|
||||
return invoice;
|
||||
// STRATEGY 2: Query each field with >>> <<< delimiters
|
||||
const [invoiceDate, invoiceDateAlt, vendor, currency, totalAmount, netAmount, vatAmount] = await Promise.all([
|
||||
queryAndLog('Invoice Date ', `Find the INVOICE DATE (when issued, NOT due date).
|
||||
Look for: "Invoice Date", "Rechnungsdatum", "Date", "Datum"
|
||||
Return ONLY the date in YYYY-MM-DD format: >>>2024-01-15<<<`),
|
||||
|
||||
// STRATEGY 3: Ask same question differently for verification
|
||||
queryAndLog('Date Alt ', `What date appears next to the invoice number at the top?
|
||||
Return ONLY YYYY-MM-DD format: >>>2024-01-15<<<`),
|
||||
|
||||
queryAndLog('Vendor ', `What company ISSUED this invoice (seller, not buyer)?
|
||||
Look at letterhead/logo at top.
|
||||
Return ONLY the company name: >>>Adobe Inc.<<<`),
|
||||
|
||||
queryAndLog('Currency ', `What currency symbol appears next to amounts? € $ or £?
|
||||
Return the 3-letter code: >>>EUR<<<`),
|
||||
|
||||
queryAndLog('Total Amount ', `What is the FINAL TOTAL amount (including tax) the customer must pay?
|
||||
Look for "Total", "Grand Total", "Gesamtbetrag" at the bottom.
|
||||
Return ONLY the number (no symbol): >>>24.99<<<`),
|
||||
|
||||
queryAndLog('Net Amount ', `What is the NET/subtotal amount BEFORE tax?
|
||||
Look for "Net", "Netto", "Subtotal".
|
||||
Return ONLY the number: >>>20.99<<<`),
|
||||
|
||||
queryAndLog('VAT Amount ', `What is the VAT/tax amount?
|
||||
Look for "VAT", "MwSt", "USt", "Tax".
|
||||
Return ONLY the number: >>>4.00<<<`),
|
||||
]);
|
||||
|
||||
// Extract value from >>> <<< delimiters, or return original if not found
|
||||
const extractDelimited = (s: string): string => {
|
||||
const match = s.match(/>>>([^<]+)<<</);
|
||||
return match ? match[1].trim() : s.trim();
|
||||
};
|
||||
|
||||
// Parse amount from string (handles European format and prose)
|
||||
const parseAmount = (s: string): number => {
|
||||
if (!s) return 0;
|
||||
|
||||
// First try delimited format
|
||||
const delimitedMatch = s.match(/>>>([^<]+)<<</);
|
||||
if (delimitedMatch) {
|
||||
const numMatch = delimitedMatch[1].match(/([\d.,]+)/);
|
||||
if (numMatch) {
|
||||
const numStr = numMatch[1];
|
||||
const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
|
||||
? numStr.replace(/\./g, '').replace(',', '.')
|
||||
: numStr.replace(/,/g, '');
|
||||
return parseFloat(normalized) || 0;
|
||||
}
|
||||
} catch (err) {
|
||||
console.log(` [Pass ${pass}] Error: ${err}`);
|
||||
}
|
||||
|
||||
// Try to find amount patterns in prose: "24.99", "24,99", "€24.99", "24.99 EUR"
|
||||
const amountPatterns = [
|
||||
/(?:€|EUR|USD|GBP)\s*([\d.,]+)/i, // €24.99 or EUR 24.99
|
||||
/([\d.,]+)\s*(?:€|EUR|USD|GBP)/i, // 24.99 EUR or 24.99€
|
||||
/(?:is|amount|total)[:\s]+([\d.,]+)/i, // "is 24.99" or "amount: 24.99"
|
||||
/\b(\d{1,3}(?:[.,]\d{2,3})*(?:[.,]\d{2}))\b/, // General number pattern with decimals
|
||||
];
|
||||
|
||||
for (const pattern of amountPatterns) {
|
||||
const match = s.match(pattern);
|
||||
if (match) {
|
||||
const numStr = match[1];
|
||||
// European format: 1.234,56 → 1234.56
|
||||
const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
|
||||
? numStr.replace(/\./g, '').replace(',', '.')
|
||||
: numStr.replace(/,/g, '');
|
||||
const value = parseFloat(normalized);
|
||||
if (value > 0) return value;
|
||||
}
|
||||
}
|
||||
|
||||
return 0;
|
||||
};
|
||||
|
||||
// STRATEGY 1: Parse "all numbers" to find invoice number
|
||||
const extractInvoiceFromList = (allNums: string): string | null => {
|
||||
const delimited = extractDelimited(allNums);
|
||||
|
||||
// Find ALL "invoice:XXX" matches
|
||||
const invoiceMatches = delimited.matchAll(/invoice[:\s]*([A-Z0-9-]+)/gi);
|
||||
const candidates: string[] = [];
|
||||
for (const match of invoiceMatches) {
|
||||
const value = match[1];
|
||||
// Filter out labels like "USt-IdNr", "INVOICE", short strings
|
||||
if (value.length > 5 && /\d{4,}/.test(value) && !/^(ust|vat|tax|nr|id|no)/i.test(value)) {
|
||||
candidates.push(value);
|
||||
}
|
||||
}
|
||||
if (candidates.length > 0) return candidates[0];
|
||||
|
||||
// Look for "rechnungsnr:XXX" pattern
|
||||
const rechnungMatch = delimited.match(/rechnung[snr]*[:\s]*([A-Z0-9-]{6,})/i);
|
||||
if (rechnungMatch && /\d{4,}/.test(rechnungMatch[1])) return rechnungMatch[1];
|
||||
|
||||
// Look for patterns like IEE2022..., R001... (Adobe invoice number patterns)
|
||||
const adobeMatch = delimited.match(/\b(IEE\d{10,})\b/i);
|
||||
if (adobeMatch) return adobeMatch[1];
|
||||
const rInvoiceMatch = delimited.match(/\b(R\d{8,})\b/i);
|
||||
if (rInvoiceMatch) return rInvoiceMatch[1];
|
||||
|
||||
return null;
|
||||
};
|
||||
|
||||
// Fallback invoice number extraction
|
||||
const extractInvoiceNumber = (s: string): string => {
|
||||
const delimited = extractDelimited(s);
|
||||
if (delimited !== s.trim()) return delimited;
|
||||
|
||||
let clean = s.replace(/\*\*/g, '').replace(/`/g, '');
|
||||
const patterns = [
|
||||
/\b([A-Z]{2,3}\d{10,})\b/i,
|
||||
/\b([A-Z]\d{8,})\b/i,
|
||||
/\b(INV[-\s]?\d{4}[-\s]?\d+)\b/i,
|
||||
/\b(\d{7,})\b/,
|
||||
];
|
||||
for (const pattern of patterns) {
|
||||
const match = clean.match(pattern);
|
||||
if (match) return match[1];
|
||||
}
|
||||
return clean.replace(/[^A-Z0-9-]/gi, '').trim() || clean.trim();
|
||||
};
|
||||
|
||||
// Extract date with fallback
|
||||
const extractDate = (s: string): string => {
|
||||
const delimited = extractDelimited(s);
|
||||
if (/^\d{4}-\d{2}-\d{2}$/.test(delimited)) return delimited;
|
||||
|
||||
let clean = s.replace(/\*\*/g, '').replace(/`/g, '');
|
||||
const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
|
||||
if (isoMatch) return isoMatch[1];
|
||||
const dmmyMatch = clean.match(/(\d{1,2})[-\/]([A-Z]{3})[-\/](\d{4})/i);
|
||||
if (dmmyMatch) {
|
||||
const monthMap: Record<string, string> = {
|
||||
JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
|
||||
JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
|
||||
};
|
||||
return `${dmmyMatch[3]}-${monthMap[dmmyMatch[2].toUpperCase()] || '01'}-${dmmyMatch[1].padStart(2, '0')}`;
|
||||
}
|
||||
const dmyMatch = clean.match(/(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})/);
|
||||
if (dmyMatch) {
|
||||
return `${dmyMatch[3]}-${dmyMatch[2].padStart(2, '0')}-${dmyMatch[1].padStart(2, '0')}`;
|
||||
}
|
||||
return '';
|
||||
};
|
||||
|
||||
// Extract currency
|
||||
const extractCurrency = (s: string): string => {
|
||||
const delimited = extractDelimited(s);
|
||||
if (['EUR', 'USD', 'GBP'].includes(delimited.toUpperCase())) return delimited.toUpperCase();
|
||||
const upper = s.toUpperCase();
|
||||
if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
|
||||
if (upper.includes('USD') || upper.includes('$')) return 'USD';
|
||||
if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
|
||||
return 'EUR';
|
||||
};
|
||||
|
||||
// Extract vendor
|
||||
const extractVendor = (s: string): string => {
|
||||
const delimited = extractDelimited(s);
|
||||
if (delimited !== s.trim()) return delimited;
|
||||
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
|
||||
if (clean.length < 50) return clean.replace(/[."]+$/, '').trim();
|
||||
const companyMatch = clean.match(/([A-Z][A-Za-z0-9\s&]+(?:Ltd|Limited|GmbH|Inc|BV|AG|SE|LLC|Co|Corp)[.]?)/i);
|
||||
if (companyMatch) return companyMatch[1].trim();
|
||||
return clean;
|
||||
};
|
||||
|
||||
// STRATEGY 1: Get invoice number - try multiple approaches
|
||||
// 1. From list with type labels
|
||||
// 2. From direct query
|
||||
// 3. From pattern matching
|
||||
const fromList = extractInvoiceFromList(allNumbers);
|
||||
const fromDirect = extractInvoiceNumber(directInvoiceNum);
|
||||
const fromFallback = extractInvoiceNumber(allNumbers);
|
||||
|
||||
// Prefer direct query if it has digits, otherwise use list
|
||||
const invoiceNumber = (fromDirect && /\d{6,}/.test(fromDirect)) ? fromDirect :
|
||||
(fromList && /\d{4,}/.test(fromList)) ? fromList :
|
||||
fromDirect || fromList || fromFallback;
|
||||
console.log(` [Parsed] Invoice Number: "${invoiceNumber}" (list: ${fromList}, direct: ${fromDirect})`);
|
||||
|
||||
// STRATEGY 3: Compare two date responses, pick the valid one
|
||||
const date1 = extractDate(invoiceDate);
|
||||
const date2 = extractDate(invoiceDateAlt);
|
||||
const finalDate = date1 || date2;
|
||||
if (date1 && date2 && date1 !== date2) {
|
||||
console.log(` [Validate] Date mismatch: "${date1}" vs "${date2}" - using first`);
|
||||
}
|
||||
|
||||
// Parse amounts
|
||||
let total = parseAmount(totalAmount);
|
||||
let net = parseAmount(netAmount);
|
||||
let vat = parseAmount(vatAmount);
|
||||
|
||||
// STRATEGY 4: Cross-field validation for amounts
|
||||
// If amounts seem wrong (e.g., 1690 instead of 1.69), try to fix
|
||||
if (total > 10000 && net < 100) {
|
||||
console.log(` [Validate] Total ${total} seems too high vs net ${net}, dividing by 100`);
|
||||
total = total / 100;
|
||||
}
|
||||
if (net > 10000 && total < 100) {
|
||||
console.log(` [Validate] Net ${net} seems too high vs total ${total}, dividing by 100`);
|
||||
net = net / 100;
|
||||
}
|
||||
|
||||
// Check if Net + VAT ≈ Total
|
||||
if (net > 0 && vat >= 0 && total > 0) {
|
||||
const calculated = net + vat;
|
||||
if (Math.abs(calculated - total) > 1) {
|
||||
console.log(` [Validate] Math check: ${net} + ${vat} = ${calculated} ≠ ${total}`);
|
||||
}
|
||||
}
|
||||
|
||||
// No consensus reached - return the most common result
|
||||
let bestHash = '';
|
||||
let bestCount = 0;
|
||||
for (const [hash, count] of hashCounts) {
|
||||
if (count > bestCount) {
|
||||
bestCount = count;
|
||||
bestHash = hash;
|
||||
}
|
||||
return {
|
||||
invoice_number: invoiceNumber,
|
||||
invoice_date: finalDate,
|
||||
vendor_name: extractVendor(vendor),
|
||||
currency: extractCurrency(currency),
|
||||
net_amount: net,
|
||||
vat_amount: vat,
|
||||
total_amount: total,
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* Normalize date to YYYY-MM-DD
|
||||
*/
|
||||
function normalizeDate(dateStr: string | null): string {
|
||||
if (!dateStr) return '';
|
||||
if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
|
||||
|
||||
const monthMap: Record<string, string> = {
|
||||
JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
|
||||
JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
|
||||
};
|
||||
|
||||
let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
|
||||
if (match) {
|
||||
return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
if (!bestHash) {
|
||||
throw new Error(`No valid results for ${invoiceName}`);
|
||||
match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
|
||||
if (match) {
|
||||
return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
const best = results.find((r) => r.hash === bestHash)!;
|
||||
console.log(` [No consensus] Using most common result (${bestCount}/${maxPasses} passes)`);
|
||||
return best.invoice;
|
||||
return dateStr;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -210,7 +378,7 @@ function compareInvoice(
|
||||
}
|
||||
|
||||
// Compare date
|
||||
if (extracted.invoice_date !== expected.invoice_date) {
|
||||
if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
|
||||
errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
|
||||
}
|
||||
|
||||
@@ -252,9 +420,7 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
|
||||
}
|
||||
}
|
||||
|
||||
// Sort alphabetically
|
||||
testCases.sort((a, b) => a.name.localeCompare(b.name));
|
||||
|
||||
return testCases;
|
||||
}
|
||||
|
||||
@@ -262,24 +428,20 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
|
||||
|
||||
tap.test('setup: ensure Docker containers are running', async () => {
|
||||
console.log('\n[Setup] Checking Docker containers...\n');
|
||||
|
||||
// Ensure MiniCPM is running
|
||||
const minicpmOk = await ensureMiniCpm();
|
||||
expect(minicpmOk).toBeTrue();
|
||||
|
||||
console.log('\n[Setup] All containers ready!\n');
|
||||
});
|
||||
|
||||
tap.test('should have MiniCPM-V 4.5 model loaded', async () => {
|
||||
tap.test('should have MiniCPM-V model loaded', async () => {
|
||||
const response = await fetch(`${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} invoice test cases (MiniCPM-V only)\n`);
|
||||
console.log(`\nFound ${testCases.length} invoice test cases (MiniCPM-V)\n`);
|
||||
|
||||
let passedCount = 0;
|
||||
let failedCount = 0;
|
||||
@@ -287,25 +449,20 @@ const processingTimes: number[] = [];
|
||||
|
||||
for (const testCase of testCases) {
|
||||
tap.test(`should extract invoice: ${testCase.name}`, async () => {
|
||||
// Load expected data
|
||||
const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
|
||||
console.log(`\n=== ${testCase.name} ===`);
|
||||
console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
|
||||
|
||||
const startTime = Date.now();
|
||||
|
||||
// Convert PDF to images
|
||||
const images = convertPdfToImages(testCase.pdfPath);
|
||||
console.log(` Pages: ${images.length}`);
|
||||
|
||||
// Extract with consensus voting (MiniCPM-V only)
|
||||
const extracted = await extractWithConsensus(images, testCase.name);
|
||||
const extracted = await extractInvoiceFromImages(images);
|
||||
console.log(` Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
|
||||
|
||||
const endTime = Date.now();
|
||||
const elapsedMs = endTime - startTime;
|
||||
const elapsedMs = Date.now() - startTime;
|
||||
processingTimes.push(elapsedMs);
|
||||
|
||||
// Compare results
|
||||
const result = compareInvoice(extracted, expected);
|
||||
|
||||
if (result.match) {
|
||||
@@ -317,7 +474,6 @@ for (const testCase of testCases) {
|
||||
result.errors.forEach((e) => console.log(` - ${e}`));
|
||||
}
|
||||
|
||||
// Assert match
|
||||
expect(result.match).toBeTrue();
|
||||
});
|
||||
}
|
||||
@@ -326,18 +482,17 @@ tap.test('summary', async () => {
|
||||
const totalInvoices = testCases.length;
|
||||
const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
|
||||
const totalTimeMs = processingTimes.reduce((a, b) => a + b, 0);
|
||||
const avgTimeMs = processingTimes.length > 0 ? totalTimeMs / processingTimes.length : 0;
|
||||
const avgTimeSec = avgTimeMs / 1000;
|
||||
const totalTimeSec = totalTimeMs / 1000;
|
||||
const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
|
||||
|
||||
console.log(`\n========================================`);
|
||||
console.log(` Invoice Extraction Summary (MiniCPM)`);
|
||||
console.log(`========================================`);
|
||||
console.log(` Method: Multi-query (no_think)`);
|
||||
console.log(` Passed: ${passedCount}/${totalInvoices}`);
|
||||
console.log(` Failed: ${failedCount}/${totalInvoices}`);
|
||||
console.log(` Accuracy: ${accuracy.toFixed(1)}%`);
|
||||
console.log(`----------------------------------------`);
|
||||
console.log(` Total time: ${totalTimeSec.toFixed(1)}s`);
|
||||
console.log(` Total time: ${(totalTimeMs / 1000).toFixed(1)}s`);
|
||||
console.log(` Avg per inv: ${avgTimeSec.toFixed(1)}s`);
|
||||
console.log(`========================================\n`);
|
||||
});
|
||||
|
||||
334
test/test.invoices.ministral3.ts
Normal file
334
test/test.invoices.ministral3.ts
Normal file
@@ -0,0 +1,334 @@
|
||||
/**
|
||||
* Invoice extraction using Ministral 3 Vision (Direct)
|
||||
*
|
||||
* NO PaddleOCR needed - Ministral 3 has built-in vision encoder:
|
||||
* 1. Convert PDF to images
|
||||
* 2. Send images directly to Ministral 3 via Ollama
|
||||
* 3. Extract structured JSON with native schema support
|
||||
*
|
||||
* This is the simplest possible pipeline.
|
||||
*/
|
||||
import { tap, expect } from '@git.zone/tstest/tapbundle';
|
||||
import * as fs from 'fs';
|
||||
import * as path from 'path';
|
||||
import { execSync } from 'child_process';
|
||||
import * as os from 'os';
|
||||
import { ensureMinistral3 } from './helpers/docker.js';
|
||||
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
const VISION_MODEL = 'ministral-3:8b';
|
||||
|
||||
interface IInvoice {
|
||||
invoice_number: string;
|
||||
invoice_date: string;
|
||||
vendor_name: string;
|
||||
currency: string;
|
||||
net_amount: number;
|
||||
vat_amount: number;
|
||||
total_amount: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert PDF to PNG images using ImageMagick
|
||||
*/
|
||||
function convertPdfToImages(pdfPath: string): string[] {
|
||||
const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
|
||||
const outputPattern = path.join(tempDir, 'page-%d.png');
|
||||
|
||||
try {
|
||||
// High quality conversion: 300 DPI, max quality, sharpen for better OCR
|
||||
execSync(
|
||||
`convert -density 300 -quality 100 "${pdfPath}" -background white -alpha remove -sharpen 0x1 "${outputPattern}"`,
|
||||
{ stdio: 'pipe' }
|
||||
);
|
||||
|
||||
const files = fs.readdirSync(tempDir).filter((f) => 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 });
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract invoice data directly from images using Ministral 3 Vision
|
||||
*/
|
||||
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
|
||||
console.log(` [Vision] Processing ${images.length} page(s) with Ministral 3`);
|
||||
|
||||
// JSON schema for structured output
|
||||
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'],
|
||||
};
|
||||
|
||||
const prompt = `You are an expert invoice data extraction system. Carefully analyze this invoice document and extract the following fields with high precision.
|
||||
|
||||
INVOICE NUMBER:
|
||||
- Look for labels: "Invoice No", "Invoice #", "Invoice Number", "Rechnung Nr", "Rechnungsnummer", "Document No", "Bill No", "Reference"
|
||||
- Usually alphanumeric, often starts with letters (e.g., R0014359508, INV-2024-001)
|
||||
- Located near the top of the invoice
|
||||
|
||||
INVOICE DATE:
|
||||
- Look for labels: "Invoice Date", "Date", "Datum", "Rechnungsdatum", "Issue Date", "Bill Date"
|
||||
- Convert ANY date format to YYYY-MM-DD (e.g., 14/10/2021 → 2021-10-14, Oct 14, 2021 → 2021-10-14)
|
||||
- Usually near the invoice number
|
||||
|
||||
VENDOR NAME:
|
||||
- The company ISSUING the invoice (not the recipient)
|
||||
- Found in letterhead, logo area, or header - typically the largest/most prominent company name
|
||||
- Examples: "Hetzner Online GmbH", "Adobe Inc", "DigitalOcean LLC"
|
||||
|
||||
CURRENCY:
|
||||
- Detect from symbols: € = EUR, $ = USD, £ = GBP
|
||||
- Or from text: "EUR", "USD", "GBP"
|
||||
- Default to EUR if unclear
|
||||
|
||||
AMOUNTS (Critical - read carefully!):
|
||||
- total_amount: The FINAL amount due/payable - look for "Total", "Grand Total", "Amount Due", "Balance Due", "Gesamtbetrag", "Endbetrag"
|
||||
- net_amount: Subtotal BEFORE tax - look for "Subtotal", "Net", "Netto", "excl. VAT"
|
||||
- vat_amount: Tax amount - look for "VAT", "Tax", "MwSt", "USt", "19%", "20%"
|
||||
- For multi-page invoices: the FINAL totals are usually on the LAST page
|
||||
|
||||
Return ONLY valid JSON with the extracted values.`;
|
||||
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({
|
||||
model: VISION_MODEL,
|
||||
messages: [
|
||||
{
|
||||
role: 'user',
|
||||
content: prompt,
|
||||
images: images, // Send all page images
|
||||
},
|
||||
],
|
||||
format: invoiceSchema,
|
||||
stream: true,
|
||||
options: {
|
||||
num_predict: 1024,
|
||||
temperature: 0.0,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
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 decoder = new TextDecoder();
|
||||
let fullText = '';
|
||||
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
const lines = chunk.split('\n').filter((l) => l.trim());
|
||||
|
||||
for (const line of lines) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
if (json.message?.content) {
|
||||
fullText += json.message.content;
|
||||
}
|
||||
} catch {
|
||||
// Skip invalid JSON lines
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 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: ${fullText.substring(0, 200)}`);
|
||||
}
|
||||
|
||||
const parsed = JSON.parse(jsonStr.substring(startIdx, endIdx));
|
||||
|
||||
return {
|
||||
invoice_number: parsed.invoice_number || null,
|
||||
invoice_date: parsed.invoice_date || null,
|
||||
vendor_name: parsed.vendor_name || null,
|
||||
currency: parsed.currency || 'EUR',
|
||||
net_amount: parseFloat(parsed.net_amount) || 0,
|
||||
vat_amount: parseFloat(parsed.vat_amount) || 0,
|
||||
total_amount: parseFloat(parsed.total_amount) || 0,
|
||||
};
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Normalize date to YYYY-MM-DD
|
||||
*/
|
||||
function normalizeDate(dateStr: string | null): string {
|
||||
if (!dateStr) return '';
|
||||
if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
|
||||
|
||||
const monthMap: Record<string, string> = {
|
||||
JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
|
||||
JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
|
||||
};
|
||||
|
||||
let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
|
||||
if (match) {
|
||||
return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
|
||||
if (match) {
|
||||
return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
return dateStr;
|
||||
}
|
||||
|
||||
/**
|
||||
* Compare extracted vs expected
|
||||
*/
|
||||
function compareInvoice(extracted: IInvoice, expected: IInvoice): { match: boolean; errors: string[] } {
|
||||
const errors: string[] = [];
|
||||
|
||||
const extNum = extracted.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
|
||||
const expNum = expected.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
|
||||
if (extNum !== expNum) {
|
||||
errors.push(`invoice_number: expected "${expected.invoice_number}", got "${extracted.invoice_number}"`);
|
||||
}
|
||||
|
||||
if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
|
||||
errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
|
||||
}
|
||||
|
||||
if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
|
||||
errors.push(`total_amount: expected ${expected.total_amount}, got ${extracted.total_amount}`);
|
||||
}
|
||||
|
||||
if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
|
||||
errors.push(`currency: expected "${expected.currency}", got "${extracted.currency}"`);
|
||||
}
|
||||
|
||||
return { match: errors.length === 0, errors };
|
||||
}
|
||||
|
||||
/**
|
||||
* Find test cases
|
||||
*/
|
||||
function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
|
||||
const testDir = path.join(process.cwd(), '.nogit/invoices');
|
||||
if (!fs.existsSync(testDir)) return [];
|
||||
|
||||
const files = fs.readdirSync(testDir);
|
||||
const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
|
||||
|
||||
for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
|
||||
const baseName = pdf.replace('.pdf', '');
|
||||
const jsonFile = `${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 Ministral 3 is running', async () => {
|
||||
console.log('\n[Setup] Checking Ministral 3...\n');
|
||||
const ok = await ensureMinistral3();
|
||||
expect(ok).toBeTrue();
|
||||
console.log('\n[Setup] Ready!\n');
|
||||
});
|
||||
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} invoice test cases (Ministral 3 Vision Direct)\n`);
|
||||
|
||||
let passedCount = 0;
|
||||
let failedCount = 0;
|
||||
const times: number[] = [];
|
||||
|
||||
for (const testCase of testCases) {
|
||||
tap.test(`should extract invoice: ${testCase.name}`, async () => {
|
||||
const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
|
||||
console.log(`\n=== ${testCase.name} ===`);
|
||||
console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
|
||||
|
||||
const start = Date.now();
|
||||
const images = convertPdfToImages(testCase.pdfPath);
|
||||
console.log(` Pages: ${images.length}`);
|
||||
|
||||
const extracted = await extractInvoiceFromImages(images);
|
||||
console.log(` Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
|
||||
const elapsed = Date.now() - start;
|
||||
times.push(elapsed);
|
||||
|
||||
const result = compareInvoice(extracted, expected);
|
||||
|
||||
if (result.match) {
|
||||
passedCount++;
|
||||
console.log(` Result: MATCH (${(elapsed / 1000).toFixed(1)}s)`);
|
||||
} else {
|
||||
failedCount++;
|
||||
console.log(` Result: MISMATCH (${(elapsed / 1000).toFixed(1)}s)`);
|
||||
result.errors.forEach((e) => console.log(` - ${e}`));
|
||||
}
|
||||
|
||||
expect(result.match).toBeTrue();
|
||||
});
|
||||
}
|
||||
|
||||
tap.test('summary', async () => {
|
||||
const total = testCases.length;
|
||||
const accuracy = total > 0 ? (passedCount / total) * 100 : 0;
|
||||
const totalTime = times.reduce((a, b) => a + b, 0) / 1000;
|
||||
const avgTime = times.length > 0 ? totalTime / times.length : 0;
|
||||
|
||||
console.log(`\n======================================================`);
|
||||
console.log(` Invoice Extraction Summary (Ministral 3 Vision)`);
|
||||
console.log(`======================================================`);
|
||||
console.log(` Method: Ministral 3 8B Vision (Direct)`);
|
||||
console.log(` Passed: ${passedCount}/${total}`);
|
||||
console.log(` Failed: ${failedCount}/${total}`);
|
||||
console.log(` Accuracy: ${accuracy.toFixed(1)}%`);
|
||||
console.log(`------------------------------------------------------`);
|
||||
console.log(` Total time: ${totalTime.toFixed(1)}s`);
|
||||
console.log(` Avg per inv: ${avgTime.toFixed(1)}s`);
|
||||
console.log(`======================================================\n`);
|
||||
});
|
||||
|
||||
export default tap.start();
|
||||
@@ -90,61 +90,59 @@ async function parseDocument(imageBase64: string): Promise<string> {
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract invoice fields from structured HTML using Qwen2.5 (text-only model)
|
||||
* Extract invoice fields using simple direct prompt
|
||||
* The OCR output has clearly labeled fields - just ask the LLM to read them
|
||||
*/
|
||||
async function extractInvoiceFromHtml(html: string): Promise<IInvoice> {
|
||||
// Truncate if too long (HTML is more valuable per byte, allow more)
|
||||
const truncated = html.length > 16000 ? html.slice(0, 16000) : html;
|
||||
console.log(` [Extract] Processing ${truncated.length} chars of HTML`);
|
||||
// OCR output is already good - just truncate if too long
|
||||
const truncated = html.length > 32000 ? html.slice(0, 32000) : html;
|
||||
console.log(` [Extract] ${truncated.length} chars of HTML`);
|
||||
|
||||
const prompt = `You are an invoice data extractor. Extract the following fields from this HTML document (OCR output with semantic structure) and return ONLY a valid JSON object.
|
||||
|
||||
The HTML uses semantic tags:
|
||||
- <table> with <thead>/<tbody> for structured tables (invoice line items, totals)
|
||||
- <header> for document header (company info, invoice number)
|
||||
- <footer> for document footer (payment terms, legal text)
|
||||
- <section class="table-region"> for table regions
|
||||
- data-type and data-y attributes indicate block type and vertical position
|
||||
|
||||
Required fields:
|
||||
- invoice_number: The invoice/receipt/document number
|
||||
- invoice_date: Date in YYYY-MM-DD format (convert from any format)
|
||||
- vendor_name: Company that issued the invoice
|
||||
- currency: EUR, USD, GBP, etc.
|
||||
- net_amount: Amount before tax (number)
|
||||
- vat_amount: Tax/VAT amount (number, use 0 if reverse charge or not shown)
|
||||
- total_amount: Final total amount (number)
|
||||
|
||||
Example output format:
|
||||
{"invoice_number":"INV-123","invoice_date":"2022-01-28","vendor_name":"Adobe","currency":"EUR","net_amount":24.99,"vat_amount":0,"total_amount":24.99}
|
||||
|
||||
Rules:
|
||||
- Return ONLY the JSON object, no explanation or markdown
|
||||
- Use null for missing string fields
|
||||
- Use 0 for missing numeric fields
|
||||
- Convert dates to YYYY-MM-DD format (e.g., "28-JAN-2022" becomes "2022-01-28")
|
||||
- Extract numbers without currency symbols
|
||||
- Look for totals in <table> sections, especially rows with "Total", "Amount Due", "Grand Total"
|
||||
|
||||
HTML Document:
|
||||
${truncated}
|
||||
|
||||
JSON:`;
|
||||
|
||||
const payload = {
|
||||
model: TEXT_MODEL,
|
||||
prompt,
|
||||
stream: true,
|
||||
options: {
|
||||
num_predict: 512,
|
||||
temperature: 0.1,
|
||||
// JSON schema for structured output
|
||||
const invoiceSchema = {
|
||||
type: 'object',
|
||||
properties: {
|
||||
invoice_number: { type: 'string' },
|
||||
invoice_date: { type: 'string' },
|
||||
vendor_name: { type: 'string' },
|
||||
currency: { type: 'string' },
|
||||
net_amount: { type: 'number' },
|
||||
vat_amount: { type: 'number' },
|
||||
total_amount: { type: 'number' },
|
||||
},
|
||||
required: ['invoice_number', 'invoice_date', 'vendor_name', 'currency', 'net_amount', 'vat_amount', 'total_amount'],
|
||||
};
|
||||
|
||||
const response = await fetch(`${OLLAMA_URL}/api/generate`, {
|
||||
// Simple, direct prompt - the OCR output already has labeled fields
|
||||
const systemPrompt = `You read invoice HTML and extract labeled fields. Return JSON only.`;
|
||||
|
||||
const userPrompt = `Extract from this invoice HTML:
|
||||
- invoice_number: Find "Invoice no.", "Invoice #", "Invoice", "Rechnung", "Document No" and extract the value
|
||||
- invoice_date: Find "Invoice date", "Date", "Datum" and convert to YYYY-MM-DD format
|
||||
- vendor_name: The company name issuing the invoice (in header/letterhead)
|
||||
- currency: EUR, USD, or GBP (look for € $ £ symbols or text)
|
||||
- total_amount: Find "Total", "Grand Total", "Amount Due", "Gesamtbetrag" - the FINAL total amount
|
||||
- net_amount: Amount before VAT/tax (Subtotal, Net)
|
||||
- vat_amount: VAT/tax amount
|
||||
|
||||
HTML:
|
||||
${truncated}
|
||||
|
||||
Return ONLY valid JSON: {"invoice_number":"...", "invoice_date":"YYYY-MM-DD", "vendor_name":"...", "currency":"EUR", "net_amount":0, "vat_amount":0, "total_amount":0}`;
|
||||
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify(payload),
|
||||
body: JSON.stringify({
|
||||
model: TEXT_MODEL,
|
||||
messages: [
|
||||
{ role: 'system', content: systemPrompt },
|
||||
{ role: 'user', content: userPrompt },
|
||||
],
|
||||
format: invoiceSchema,
|
||||
stream: true,
|
||||
options: { num_predict: 512, temperature: 0.0 },
|
||||
}),
|
||||
});
|
||||
|
||||
if (!response.ok) {
|
||||
@@ -169,7 +167,9 @@ JSON:`;
|
||||
for (const line of lines) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
if (json.response) {
|
||||
if (json.message?.content) {
|
||||
fullText += json.message.content;
|
||||
} else if (json.response) {
|
||||
fullText += json.response;
|
||||
}
|
||||
} catch {
|
||||
@@ -179,17 +179,37 @@ JSON:`;
|
||||
}
|
||||
|
||||
// Extract JSON from response
|
||||
const startIdx = fullText.indexOf('{');
|
||||
const endIdx = fullText.lastIndexOf('}') + 1;
|
||||
let jsonStr = fullText.trim();
|
||||
|
||||
// Remove markdown code block if present
|
||||
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();
|
||||
|
||||
// Find JSON object boundaries
|
||||
const startIdx = jsonStr.indexOf('{');
|
||||
const endIdx = jsonStr.lastIndexOf('}') + 1;
|
||||
|
||||
if (startIdx < 0 || endIdx <= startIdx) {
|
||||
throw new Error(`No JSON object found in response: ${fullText.substring(0, 200)}`);
|
||||
}
|
||||
|
||||
const jsonStr = fullText.substring(startIdx, endIdx);
|
||||
const parsed = JSON.parse(jsonStr);
|
||||
jsonStr = jsonStr.substring(startIdx, endIdx);
|
||||
|
||||
// Ensure numeric fields are actually numbers
|
||||
let parsed;
|
||||
try {
|
||||
parsed = JSON.parse(jsonStr);
|
||||
} catch (e) {
|
||||
throw new Error(`Invalid JSON: ${jsonStr.substring(0, 200)}`);
|
||||
}
|
||||
|
||||
// Normalize response to expected format
|
||||
return {
|
||||
invoice_number: parsed.invoice_number || null,
|
||||
invoice_date: parsed.invoice_date || null,
|
||||
@@ -203,14 +223,23 @@ JSON:`;
|
||||
|
||||
/**
|
||||
* Single extraction pass: Parse with PaddleOCR-VL Full, extract with Qwen2.5 (text-only)
|
||||
* Processes ALL pages and concatenates HTML for multi-page invoice support
|
||||
*/
|
||||
async function extractOnce(images: string[], passNum: number): Promise<IInvoice> {
|
||||
// Parse document with full pipeline (PaddleOCR-VL) -> returns HTML
|
||||
const html = await parseDocument(images[0]);
|
||||
console.log(` [Parse] Got ${html.split('\n').length} lines of HTML`);
|
||||
// Parse ALL pages and concatenate HTML with page markers
|
||||
const htmlParts: string[] = [];
|
||||
|
||||
for (let i = 0; i < images.length; i++) {
|
||||
const pageHtml = await parseDocument(images[i]);
|
||||
// Add page marker for context
|
||||
htmlParts.push(`<!-- Page ${i + 1} -->\n${pageHtml}`);
|
||||
}
|
||||
|
||||
const fullHtml = htmlParts.join('\n\n');
|
||||
console.log(` [Parse] Got ${fullHtml.split('\n').length} lines from ${images.length} page(s)`);
|
||||
|
||||
// Extract invoice fields from HTML using text-only model (no images)
|
||||
return extractInvoiceFromHtml(html);
|
||||
return extractInvoiceFromHtml(fullHtml);
|
||||
}
|
||||
|
||||
/**
|
||||
|
||||
351
test/test.invoices.qwen3vl.ts
Normal file
351
test/test.invoices.qwen3vl.ts
Normal file
@@ -0,0 +1,351 @@
|
||||
/**
|
||||
* Invoice extraction using Qwen3-VL 8B Vision (Direct)
|
||||
*
|
||||
* Multi-query approach: 5 parallel simple queries to avoid token exhaustion.
|
||||
* Single pass, no consensus voting.
|
||||
*/
|
||||
import { tap, expect } from '@git.zone/tstest/tapbundle';
|
||||
import * as fs from 'fs';
|
||||
import * as path from 'path';
|
||||
import { execSync } from 'child_process';
|
||||
import * as os from 'os';
|
||||
import { ensureMiniCpm } from './helpers/docker.js';
|
||||
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
const VISION_MODEL = 'qwen3-vl:8b';
|
||||
|
||||
interface IInvoice {
|
||||
invoice_number: string;
|
||||
invoice_date: string;
|
||||
vendor_name: string;
|
||||
currency: string;
|
||||
net_amount: number;
|
||||
vat_amount: number;
|
||||
total_amount: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert PDF to PNG images using ImageMagick
|
||||
*/
|
||||
function convertPdfToImages(pdfPath: string): string[] {
|
||||
const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pdf-convert-'));
|
||||
const outputPattern = path.join(tempDir, 'page-%d.png');
|
||||
|
||||
try {
|
||||
// 150 DPI is sufficient for invoice extraction, reduces context size
|
||||
execSync(
|
||||
`convert -density 150 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
|
||||
{ stdio: 'pipe' }
|
||||
);
|
||||
|
||||
const files = fs.readdirSync(tempDir).filter((f) => f.endsWith('.png')).sort();
|
||||
const images: string[] = [];
|
||||
|
||||
for (const file of files) {
|
||||
const imagePath = path.join(tempDir, file);
|
||||
const imageData = fs.readFileSync(imagePath);
|
||||
images.push(imageData.toString('base64'));
|
||||
}
|
||||
|
||||
return images;
|
||||
} finally {
|
||||
fs.rmSync(tempDir, { recursive: true, force: true });
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Query Qwen3-VL for a single field
|
||||
* Uses simple prompts to minimize thinking tokens
|
||||
*/
|
||||
async function queryField(images: string[], question: string): Promise<string> {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({
|
||||
model: VISION_MODEL,
|
||||
messages: [{
|
||||
role: 'user',
|
||||
content: `${question} Reply with just the value, nothing else.`,
|
||||
images: images,
|
||||
}],
|
||||
stream: false,
|
||||
options: {
|
||||
num_predict: 500,
|
||||
temperature: 0.1,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
if (!response.ok) {
|
||||
throw new Error(`Ollama API error: ${response.status}`);
|
||||
}
|
||||
|
||||
const data = await response.json();
|
||||
return (data.message?.content || '').trim();
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract invoice data using multiple simple queries
|
||||
* Each query asks for 1-2 fields to minimize thinking tokens
|
||||
* (Qwen3's thinking mode uses all tokens on complex prompts)
|
||||
*/
|
||||
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
|
||||
console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`);
|
||||
|
||||
// Query each field separately to avoid excessive thinking tokens
|
||||
// Use explicit questions to avoid confusion between similar fields
|
||||
// Log each result as it comes in (not waiting for all to complete)
|
||||
const queryAndLog = async (name: string, question: string): Promise<string> => {
|
||||
const result = await queryField(images, question);
|
||||
console.log(` [Query] ${name}: "${result}"`);
|
||||
return result;
|
||||
};
|
||||
|
||||
const [invoiceNum, invoiceDate, vendor, currency, totalAmount, netAmount, vatAmount] = await Promise.all([
|
||||
queryAndLog('Invoice Number', 'What is the INVOICE NUMBER (not VAT number, not customer ID)? Look for "Invoice No", "Invoice #", "Rechnung Nr", "Facture". Just the number/code.'),
|
||||
queryAndLog('Invoice Date ', 'What is the INVOICE DATE (not due date, not delivery date)? The date the invoice was issued. Format: YYYY-MM-DD'),
|
||||
queryAndLog('Vendor ', 'What company ISSUED this invoice (the seller/vendor, not the buyer)? Look at the letterhead or "From" section.'),
|
||||
queryAndLog('Currency ', 'What CURRENCY is used? Look for € (EUR), $ (USD), or £ (GBP). Answer with 3-letter code: EUR, USD, or GBP'),
|
||||
queryAndLog('Total Amount ', 'What is the TOTAL AMOUNT INCLUDING TAX (the final amount to pay, with VAT/tax included)? Just the number, e.g. 24.99'),
|
||||
queryAndLog('Net Amount ', 'What is the NET AMOUNT (subtotal before VAT/tax)? Just the number, e.g. 20.99'),
|
||||
queryAndLog('VAT Amount ', 'What is the VAT/TAX AMOUNT? Just the number, e.g. 4.00'),
|
||||
]);
|
||||
|
||||
// Parse amount from string (handles European format)
|
||||
const parseAmount = (s: string): number => {
|
||||
if (!s) return 0;
|
||||
// Extract number from the response
|
||||
const match = s.match(/([\d.,]+)/);
|
||||
if (!match) return 0;
|
||||
const numStr = match[1];
|
||||
// Handle European format: 1.234,56 → 1234.56
|
||||
const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
|
||||
? numStr.replace(/\./g, '').replace(',', '.')
|
||||
: numStr.replace(/,/g, '');
|
||||
return parseFloat(normalized) || 0;
|
||||
};
|
||||
|
||||
// Extract invoice number from potentially verbose response
|
||||
const extractInvoiceNumber = (s: string): string => {
|
||||
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
|
||||
// Look for common invoice number patterns
|
||||
const patterns = [
|
||||
/\b([A-Z]{2,3}\d{10,})\b/i, // IEE2022006460244
|
||||
/\b([A-Z]\d{8,})\b/i, // R0014359508
|
||||
/\b(INV[-\s]?\d{4}[-\s]?\d+)\b/i, // INV-2024-001
|
||||
/\b(\d{7,})\b/, // 1579087430
|
||||
];
|
||||
for (const pattern of patterns) {
|
||||
const match = clean.match(pattern);
|
||||
if (match) return match[1];
|
||||
}
|
||||
return clean.replace(/[^A-Z0-9-]/gi, '').trim() || clean;
|
||||
};
|
||||
|
||||
// Extract date (YYYY-MM-DD) from response
|
||||
const extractDate = (s: string): string => {
|
||||
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
|
||||
const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
|
||||
if (isoMatch) return isoMatch[1];
|
||||
return clean.replace(/[^\d-]/g, '').trim();
|
||||
};
|
||||
|
||||
// Extract currency
|
||||
const extractCurrency = (s: string): string => {
|
||||
const upper = s.toUpperCase();
|
||||
if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
|
||||
if (upper.includes('USD') || upper.includes('$')) return 'USD';
|
||||
if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
|
||||
return 'EUR';
|
||||
};
|
||||
|
||||
return {
|
||||
invoice_number: extractInvoiceNumber(invoiceNum),
|
||||
invoice_date: extractDate(invoiceDate),
|
||||
vendor_name: vendor.replace(/\*\*/g, '').replace(/`/g, '').trim() || '',
|
||||
currency: extractCurrency(currency),
|
||||
net_amount: parseAmount(netAmount),
|
||||
vat_amount: parseAmount(vatAmount),
|
||||
total_amount: parseAmount(totalAmount),
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* Normalize date to YYYY-MM-DD
|
||||
*/
|
||||
function normalizeDate(dateStr: string | null): string {
|
||||
if (!dateStr) return '';
|
||||
if (/^\d{4}-\d{2}-\d{2}$/.test(dateStr)) return dateStr;
|
||||
|
||||
const monthMap: Record<string, string> = {
|
||||
JAN: '01', FEB: '02', MAR: '03', APR: '04', MAY: '05', JUN: '06',
|
||||
JUL: '07', AUG: '08', SEP: '09', OCT: '10', NOV: '11', DEC: '12',
|
||||
};
|
||||
|
||||
let match = dateStr.match(/^(\d{1,2})-([A-Z]{3})-(\d{4})$/i);
|
||||
if (match) {
|
||||
return `${match[3]}-${monthMap[match[2].toUpperCase()] || '01'}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
|
||||
if (match) {
|
||||
return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
|
||||
}
|
||||
|
||||
return dateStr;
|
||||
}
|
||||
|
||||
/**
|
||||
* Compare extracted vs expected
|
||||
*/
|
||||
function compareInvoice(extracted: IInvoice, expected: IInvoice): { match: boolean; errors: string[] } {
|
||||
const errors: string[] = [];
|
||||
|
||||
const extNum = extracted.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
|
||||
const expNum = expected.invoice_number?.replace(/\s/g, '').toLowerCase() || '';
|
||||
if (extNum !== expNum) {
|
||||
errors.push(`invoice_number: expected "${expected.invoice_number}", got "${extracted.invoice_number}"`);
|
||||
}
|
||||
|
||||
if (normalizeDate(extracted.invoice_date) !== normalizeDate(expected.invoice_date)) {
|
||||
errors.push(`invoice_date: expected "${expected.invoice_date}", got "${extracted.invoice_date}"`);
|
||||
}
|
||||
|
||||
if (Math.abs(extracted.total_amount - expected.total_amount) > 0.02) {
|
||||
errors.push(`total_amount: expected ${expected.total_amount}, got ${extracted.total_amount}`);
|
||||
}
|
||||
|
||||
if (extracted.currency?.toUpperCase() !== expected.currency?.toUpperCase()) {
|
||||
errors.push(`currency: expected "${expected.currency}", got "${extracted.currency}"`);
|
||||
}
|
||||
|
||||
return { match: errors.length === 0, errors };
|
||||
}
|
||||
|
||||
/**
|
||||
* Find test cases
|
||||
*/
|
||||
function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: string }> {
|
||||
const testDir = path.join(process.cwd(), '.nogit/invoices');
|
||||
if (!fs.existsSync(testDir)) return [];
|
||||
|
||||
const files = fs.readdirSync(testDir);
|
||||
const testCases: Array<{ name: string; pdfPath: string; jsonPath: string }> = [];
|
||||
|
||||
for (const pdf of files.filter((f) => f.endsWith('.pdf'))) {
|
||||
const baseName = pdf.replace('.pdf', '');
|
||||
const jsonFile = `${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));
|
||||
}
|
||||
|
||||
/**
|
||||
* 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...\n');
|
||||
|
||||
// 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');
|
||||
});
|
||||
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} invoice test cases (Qwen3-VL Vision)\n`);
|
||||
|
||||
let passedCount = 0;
|
||||
let failedCount = 0;
|
||||
const times: number[] = [];
|
||||
|
||||
for (const testCase of testCases) {
|
||||
tap.test(`should extract invoice: ${testCase.name}`, async () => {
|
||||
const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
|
||||
console.log(`\n=== ${testCase.name} ===`);
|
||||
console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
|
||||
|
||||
const start = Date.now();
|
||||
const images = convertPdfToImages(testCase.pdfPath);
|
||||
console.log(` Pages: ${images.length}`);
|
||||
|
||||
const extracted = await extractInvoiceFromImages(images);
|
||||
console.log(` Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
|
||||
const elapsed = Date.now() - start;
|
||||
times.push(elapsed);
|
||||
|
||||
const result = compareInvoice(extracted, expected);
|
||||
|
||||
if (result.match) {
|
||||
passedCount++;
|
||||
console.log(` Result: MATCH (${(elapsed / 1000).toFixed(1)}s)`);
|
||||
} else {
|
||||
failedCount++;
|
||||
console.log(` Result: MISMATCH (${(elapsed / 1000).toFixed(1)}s)`);
|
||||
result.errors.forEach((e) => console.log(` - ${e}`));
|
||||
}
|
||||
|
||||
expect(result.match).toBeTrue();
|
||||
});
|
||||
}
|
||||
|
||||
tap.test('summary', async () => {
|
||||
const total = testCases.length;
|
||||
const accuracy = total > 0 ? (passedCount / total) * 100 : 0;
|
||||
const totalTime = times.reduce((a, b) => a + b, 0) / 1000;
|
||||
const avgTime = times.length > 0 ? totalTime / times.length : 0;
|
||||
|
||||
console.log(`\n======================================================`);
|
||||
console.log(` Invoice Extraction Summary (Qwen3-VL Vision)`);
|
||||
console.log(`======================================================`);
|
||||
console.log(` Method: Multi-query (single pass)`);
|
||||
console.log(` Passed: ${passedCount}/${total}`);
|
||||
console.log(` Failed: ${failedCount}/${total}`);
|
||||
console.log(` Accuracy: ${accuracy.toFixed(1)}%`);
|
||||
console.log(`------------------------------------------------------`);
|
||||
console.log(` Total time: ${totalTime.toFixed(1)}s`);
|
||||
console.log(` Avg per inv: ${avgTime.toFixed(1)}s`);
|
||||
console.log(`======================================================\n`);
|
||||
});
|
||||
|
||||
export default tap.start();
|
||||
Reference in New Issue
Block a user