host.today/ht-docker-ai
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Jürgen Kunz
2026-01-18 11:26:38 +00:00
parent 4c368dfef9
commit 76b21f1f7b
5 changed files with 624 additions and 367 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
changelog.md
View File

@@ -1,5 +1,14 @@
# Changelog # 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) ## 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 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

296
test/test.bankstatements.minicpm.ts
View File

@@ -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 * Multi-query approach with thinking DISABLED for speed:
* without any OCR augmentation. * 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 { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs'; import * as fs from 'fs';
@@ -11,24 +13,8 @@ import { execSync } from 'child_process';
import * as os from 'os'; import * as os from 'os';
import { ensureMiniCpm } from './helpers/docker.js'; import { ensureMiniCpm } from './helpers/docker.js';
// Service URL
const OLLAMA_URL = 'http://localhost:11434'; const OLLAMA_URL = 'http://localhost:11434';
const MODEL = 'minicpm-v:latest';
// 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.`;
interface ITransaction { interface ITransaction {
date: string; 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[]> { async function queryVision(image: string, prompt: string): Promise<string> {
const payload = {
model: MINICPM_MODEL,
prompt: MINICPM_EXTRACT_PROMPT,
images,
stream: true,
options: {
num_predict: 16384,
temperature: 0.1,
},
};
const response = await fetch(`${OLLAMA_URL}/api/generate`, { const response = await fetch(`${OLLAMA_URL}/api/generate`, {
method: 'POST', method: 'POST',
headers: { 'Content-Type': 'application/json' }, 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) { if (!response.ok) {
throw new Error(`Ollama API error: ${response.status}`); throw new Error(`Ollama API error: ${response.status}`);
} }
const reader = response.body?.getReader(); const data = await response.json();
if (!reader) { return (data.response || '').trim();
throw new Error('No response body'); }
}
const decoder = new TextDecoder(); /**
let fullText = ''; * Count transactions on a page
let lineBuffer = ''; */
async function countTransactions(image: string, pageNum: number): Promise<number> {
const response = await queryVision(image,
`Count the transaction rows in this bank statement table.
Each transaction has a date, description, and amount (debit or credit).
Do not count headers or totals.
How many transaction rows are there? Answer with just the number.`
);
console.log(`[${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: ${count}`);
return count;
}
while (true) { /**
const { done, value } = await reader.read(); * Get a single transaction by index (logs immediately)
if (done) break; */
async function getTransaction(image: string, index: number, pageNum: number): Promise<ITransaction | null> {
const response = await queryVision(image,
`Look at transaction row #${index} in the bank statement table (row 1 is the first transaction after the header).
const chunk = decoder.decode(value, { stream: true }); Extract:
const lines = chunk.split('\n').filter((l) => l.trim()); - 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) { Format your answer as: DATE|COUNTERPARTY|AMOUNT
try { Example: 2024-01-15|Amazon|-25.99`
const json = JSON.parse(line); );
if (json.response) {
fullText += json.response;
lineBuffer += json.response;
if (lineBuffer.includes('\n')) { // Parse the response
const parts = lineBuffer.split('\n'); const lines = response.split('\n').filter(l => l.includes('|'));
for (let i = 0; i < parts.length - 1; i++) { const line = lines[lines.length - 1] || response;
console.log(parts[i]); const parts = line.split('|').map(p => p.trim());
}
lineBuffer = parts[parts.length - 1]; if (parts.length >= 3) {
} // Parse amount - handle various formats
} let amountStr = parts[2].replace(/[€$£\s]/g, '').replace('−', '-').replace('–', '-');
} catch { // European format: comma is decimal
// Skip invalid JSON lines if (amountStr.includes(',')) {
} amountStr = amountStr.replace(/\./g, '').replace(',', '.');
} }
const amount = parseFloat(amountStr) || 0;
const tx = {
date: parts[0],
counterparty: parts[1],
amount: amount,
};
// Log immediately as this transaction completes
console.log(` [P${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) { // Log raw response on parse failure
console.log(lineBuffer); console.log(` [P${pageNum} Tx${index.toString().padStart(2, ' ')}] PARSE FAILED: "${response.replace(/\n/g, ' ').substring(0, 60)}..."`);
} return null;
console.log('');
const startIdx = fullText.indexOf('[');
const endIdx = fullText.lastIndexOf(']') + 1;
if (startIdx < 0 || endIdx <= startIdx) {
throw new Error('No JSON array found in response');
}
return JSON.parse(fullText.substring(startIdx, endIdx));
} }
/** /**
* Create a hash of transactions for comparison * Extract transactions from a single page using multi-query approach
*/ */
function hashTransactions(transactions: ITransaction[]): string { async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> {
return transactions // Step 1: Count transactions
.map((t) => `${t.date}|${t.amount.toFixed(2)}`) const count = await countTransactions(image, pageNum);
.sort()
.join(';');
}
/** if (count === 0) {
* Extract with consensus voting using MiniCPM-V only return [];
*/ }
async function extractWithConsensus(
images: string[],
maxPasses: number = 5
): Promise<ITransaction[]> {
const results: Array<{ transactions: ITransaction[]; hash: string }> = [];
const hashCounts: Map<string, number> = new Map();
const addResult = (transactions: ITransaction[], passLabel: string): number => { // Step 2: Query each transaction (in batches to avoid overwhelming)
const hash = hashTransactions(transactions); // Each transaction logs itself as it completes
results.push({ transactions, hash }); const transactions: ITransaction[] = [];
hashCounts.set(hash, (hashCounts.get(hash) || 0) + 1); const batchSize = 5;
console.log(
`[${passLabel}] Got ${transactions.length} transactions (hash: ${hash.substring(0, 20)}...)` 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 (const tx of results) {
if (tx) {
for (let pass = 1; pass <= maxPasses; pass++) { transactions.push(tx);
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;
} }
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 console.log(` [Page ${pageNum}] Complete: ${transactions.length}/${count} extracted`);
let bestHash = ''; return transactions;
let bestCount = 0; }
for (const [hash, count] of hashCounts) {
if (count > bestCount) { /**
bestCount = count; * Extract all transactions from bank statement
bestHash = hash; */
} 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) { console.log(` [Vision] Total: ${allTransactions.length} transactions`);
throw new Error('No valid results obtained'); return allTransactions;
}
const best = results.find((r) => r.hash === bestHash)!;
console.log(`[No consensus] Using most common result (${bestCount}/${maxPasses} passes)`);
return best.transactions;
} }
/** /**
@@ -273,62 +256,69 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
} }
} }
return testCases; return testCases.sort((a, b) => a.name.localeCompare(b.name));
} }
// Tests // Tests
tap.test('setup: ensure Docker containers are running', async () => { tap.test('setup: ensure Docker containers are running', async () => {
console.log('\n[Setup] Checking Docker containers...\n'); console.log('\n[Setup] Checking Docker containers...\n');
// Ensure MiniCPM is running
const minicpmOk = await ensureMiniCpm(); const minicpmOk = await ensureMiniCpm();
expect(minicpmOk).toBeTrue(); expect(minicpmOk).toBeTrue();
console.log('\n[Setup] All containers ready!\n'); console.log('\n[Setup] All containers ready!\n');
}); });
tap.test('should have MiniCPM-V 4.5 model loaded', async () => { tap.test('should have MiniCPM-V model loaded', async () => {
const response = await fetch(`${OLLAMA_URL}/api/tags`); const response = await fetch(`${OLLAMA_URL}/api/tags`);
const data = await response.json(); const data = await response.json();
const modelNames = data.models.map((m: { name: string }) => m.name); 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(); 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) { for (const testCase of testCases) {
tap.test(`should extract transactions from ${testCase.name}`, async () => { tap.test(`should extract: ${testCase.name}`, async () => {
// Load expected transactions
const expected: ITransaction[] = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8')); const expected: ITransaction[] = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
console.log(`\n=== ${testCase.name} ===`); console.log(`\n=== ${testCase.name} ===`);
console.log(`Expected: ${expected.length} transactions`); console.log(`Expected: ${expected.length} transactions`);
// Convert PDF to images
console.log('Converting PDF to images...');
const images = convertPdfToImages(testCase.pdfPath); 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 extractTransactions(images);
const extracted = await extractWithConsensus(images); console.log(` Extracted: ${extracted.length} transactions`);
console.log(`\nFinal: ${extracted.length} transactions`);
// Compare results
const result = compareTransactions(extracted, expected); 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) { if (accuracy >= 0.95 && extracted.length === expected.length) {
console.log('Errors:'); passedCount++;
result.errors.forEach((e) => console.log(` - ${e}`)); console.log(` Result: PASS (${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(accuracy).toBeGreaterThan(0.95);
expect(extracted.length).toEqual(expected.length); 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(); export default tap.start();

163
test/test.bankstatements.qwen3vl.ts
View File

@@ -1,12 +1,10 @@
/** /**
* Bank statement extraction using Qwen3-VL 8B Vision (Direct) * Bank statement extraction using Qwen3-VL 8B Vision (Direct)
* *
* Single-step pipeline: PDF → Images → Qwen3-VL → JSON * Multi-query approach:
* * 1. First ask how many transactions on each page
* Key insights: * 2. Then query each transaction individually
* - Use /no_think in prompt + think:false in API to disable reasoning * Single pass, no consensus voting.
* - Need high num_predict (8000+) for many transactions
* - Single pass extraction, no consensus needed
*/ */
import { tap, expect } from '@git.zone/tstest/tapbundle'; import { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs'; import * as fs from 'fs';
@@ -53,15 +51,9 @@ function convertPdfToImages(pdfPath: string): string[] {
} }
/** /**
* Extract transactions from a single page * Query Qwen3-VL with a simple prompt
* Processes one page at a time to minimize thinking tokens
*/ */
async function extractTransactionsFromPage(image: string, pageNum: number): Promise<ITransaction[]> { async function queryVision(image: string, prompt: string): Promise<string> {
const prompt = `/no_think
Extract transactions from this bank statement page.
Amount: "- 21,47 €" = -21.47, "+ 1.000,00 €" = 1000.00 (European format)
Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21.47},...]`;
const response = await fetch(`${OLLAMA_URL}/api/chat`, { const response = await fetch(`${OLLAMA_URL}/api/chat`, {
method: 'POST', method: 'POST',
headers: { 'Content-Type': 'application/json' }, headers: { 'Content-Type': 'application/json' },
@@ -73,9 +65,8 @@ Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21
images: [image], images: [image],
}], }],
stream: false, stream: false,
think: false,
options: { options: {
num_predict: 4000, num_predict: 500,
temperature: 0.1, temperature: 0.1,
}, },
}), }),
@@ -86,47 +77,116 @@ Return JSON array only: [{"date":"YYYY-MM-DD","counterparty":"NAME","amount":-21
} }
const data = await response.json(); const data = await response.json();
let content = data.message?.content || ''; return (data.message?.content || '').trim();
if (!content) {
console.log(` [Page ${pageNum}] Empty response`);
return [];
}
// Parse JSON array
if (content.startsWith('```json')) content = content.slice(7);
else if (content.startsWith('```')) content = content.slice(3);
if (content.endsWith('```')) content = content.slice(0, -3);
content = content.trim();
const startIdx = content.indexOf('[');
const endIdx = content.lastIndexOf(']') + 1;
if (startIdx < 0 || endIdx <= startIdx) {
console.log(` [Page ${pageNum}] No JSON array found`);
return [];
}
try {
const transactions = JSON.parse(content.substring(startIdx, endIdx));
console.log(` [Page ${pageNum}] Found ${transactions.length} transactions`);
return transactions;
} catch {
console.log(` [Page ${pageNum}] JSON parse error`);
return [];
}
} }
/** /**
* Extract transactions using Qwen3-VL vision * Count transactions on a page
* Processes each page separately to avoid thinking token exhaustion */
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[]> { async function extractTransactions(images: string[]): Promise<ITransaction[]> {
console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL`); console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`);
const allTransactions: ITransaction[] = []; const allTransactions: ITransaction[] = [];
// Process pages sequentially to avoid overwhelming the model
for (let i = 0; i < images.length; i++) { for (let i = 0; i < images.length; i++) {
const pageTransactions = await extractTransactionsFromPage(images[i], i + 1); const pageTransactions = await extractTransactionsFromPage(images[i], i + 1);
allTransactions.push(...pageTransactions); allTransactions.push(...pageTransactions);
@@ -276,8 +336,9 @@ tap.test('summary', async () => {
console.log(`\n======================================================`); console.log(`\n======================================================`);
console.log(` Bank Statement Summary (Qwen3-VL Vision)`); console.log(` Bank Statement Summary (Qwen3-VL Vision)`);
console.log(`======================================================`); console.log(`======================================================`);
console.log(` Passed: ${passedCount}/${total}`); console.log(` Method: Multi-query (count then extract each)`);
console.log(` Failed: ${failedCount}/${total}`); console.log(` Passed: ${passedCount}/${total}`);
console.log(` Failed: ${failedCount}/${total}`);
console.log(`======================================================\n`); console.log(`======================================================\n`);
}); });

427
test/test.invoices.minicpm.ts
View File

@@ -1,8 +1,8 @@
/** /**
* Invoice extraction test using MiniCPM-V only (visual extraction) * Invoice extraction test using MiniCPM-V only (visual extraction)
* *
* This tests MiniCPM-V's ability to extract invoice data directly from images * Multi-query approach with thinking DISABLED for speed.
* without any OCR augmentation. * Single pass, no consensus voting.
*/ */
import { tap, expect } from '@git.zone/tstest/tapbundle'; import { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs'; import * as fs from 'fs';
@@ -24,28 +24,6 @@ interface IInvoice {
total_amount: number; 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 * 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> { async function queryField(images: string[], question: string): Promise<string> {
const payload = {
model: MODEL,
prompt: buildPrompt(),
images,
stream: true,
options: {
num_predict: 2048,
temperature: 0.1,
},
};
const response = await fetch(`${OLLAMA_URL}/api/generate`, { const response = await fetch(`${OLLAMA_URL}/api/generate`, {
method: 'POST', method: 'POST',
headers: { 'Content-Type': 'application/json' }, 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) { if (!response.ok) {
throw new Error(`Ollama API error: ${response.status}`); throw new Error(`Ollama API error: ${response.status}`);
} }
const reader = response.body?.getReader(); const data = await response.json();
if (!reader) { const content = (data.response || '').trim();
throw new Error('No response body');
}
const decoder = new TextDecoder(); // Return full content (no thinking to filter)
let fullText = ''; return content;
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: ${fullText.substring(0, 200)}`);
}
const jsonStr = fullText.substring(startIdx, endIdx);
return JSON.parse(jsonStr);
} }
/** /**
* Create a hash of invoice for comparison (using key fields) * Extract invoice data using multiple queries with validation
*/ */
function hashInvoice(invoice: IInvoice): string { async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
return `${invoice.invoice_number}|${invoice.invoice_date}|${invoice.total_amount.toFixed(2)}`; console.log(` [Vision] Processing ${images.length} page(s) with MiniCPM-V (multi-query + validation)`);
}
/** // Log each result as it comes in
* Extract with consensus voting using MiniCPM-V only const queryAndLog = async (name: string, question: string): Promise<string> => {
*/ const result = await queryField(images, question);
async function extractWithConsensus(images: string[], invoiceName: string, maxPasses: number = 5): Promise<IInvoice> { console.log(` [Query] ${name}: "${result}"`);
const results: Array<{ invoice: IInvoice; hash: string }> = []; return result;
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)!;
}; };
for (let pass = 1; pass <= maxPasses; pass++) { // STRATEGY 1: List-then-pick for invoice number (avoids confusion with VAT/customer IDs)
try { // Also ask for invoice number directly as backup
const invoice = await extractOnce(images, pass); const [allNumbers, directInvoiceNum] = await Promise.all([
const count = addResult(invoice, `Pass ${pass}`); 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) { // STRATEGY 2: Query each field with >>> <<< delimiters
console.log(` [Consensus] Reached after ${pass} passes`); const [invoiceDate, invoiceDateAlt, vendor, currency, totalAmount, netAmount, vatAmount] = await Promise.all([
return invoice; 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} ≠ ${total}`);
} }
} }
// No consensus reached - return the most common result return {
let bestHash = ''; invoice_number: invoiceNumber,
let bestCount = 0; invoice_date: finalDate,
for (const [hash, count] of hashCounts) { vendor_name: extractVendor(vendor),
if (count > bestCount) { currency: extractCurrency(currency),
bestCount = count; net_amount: net,
bestHash = hash; vat_amount: vat,
} total_amount: total,
};
}
/**
* Normalize date to YYYY-MM-DD
*/
function normalizeDate(dateStr: string | null): string {
if (!dateStr) return '';
if (/^\d{4}-\d{2}-\d{2}$/.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) { match = dateStr.match(/^(\d{1,2})[\/.](\d{1,2})[\/.](\d{4})$/);
throw new Error(`No valid results for ${invoiceName}`); if (match) {
return `${match[3]}-${match[2].padStart(2, '0')}-${match[1].padStart(2, '0')}`;
} }
const best = results.find((r) => r.hash === bestHash)!; return dateStr;
console.log(` [No consensus] Using most common result (${bestCount}/${maxPasses} passes)`);
return best.invoice;
} }
/** /**
@@ -210,7 +378,7 @@ function compareInvoice(
} }
// Compare date // 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}"`); 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)); testCases.sort((a, b) => a.name.localeCompare(b.name));
return testCases; return testCases;
} }
@@ -262,24 +428,20 @@ function findTestCases(): Array<{ name: string; pdfPath: string; jsonPath: strin
tap.test('setup: ensure Docker containers are running', async () => { tap.test('setup: ensure Docker containers are running', async () => {
console.log('\n[Setup] Checking Docker containers...\n'); console.log('\n[Setup] Checking Docker containers...\n');
// Ensure MiniCPM is running
const minicpmOk = await ensureMiniCpm(); const minicpmOk = await ensureMiniCpm();
expect(minicpmOk).toBeTrue(); expect(minicpmOk).toBeTrue();
console.log('\n[Setup] All containers ready!\n'); 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(`${OLLAMA_URL}/api/tags`); const response = await fetch(`${OLLAMA_URL}/api/tags`);
const data = await response.json(); const data = await response.json();
const modelNames = data.models.map((m: { name: string }) => m.name); 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(); 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 passedCount = 0;
let failedCount = 0; let failedCount = 0;
@@ -287,25 +449,20 @@ const processingTimes: number[] = [];
for (const testCase of testCases) { for (const testCase of testCases) {
tap.test(`should extract invoice: ${testCase.name}`, async () => { tap.test(`should extract invoice: ${testCase.name}`, async () => {
// Load expected data
const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8')); const expected: IInvoice = JSON.parse(fs.readFileSync(testCase.jsonPath, 'utf-8'));
console.log(`\n=== ${testCase.name} ===`); console.log(`\n=== ${testCase.name} ===`);
console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`); console.log(`Expected: ${expected.invoice_number} | ${expected.invoice_date} | ${expected.total_amount} ${expected.currency}`);
const startTime = Date.now(); const startTime = Date.now();
// Convert PDF to images
const images = convertPdfToImages(testCase.pdfPath); const images = convertPdfToImages(testCase.pdfPath);
console.log(` Pages: ${images.length}`); console.log(` Pages: ${images.length}`);
// Extract with consensus voting (MiniCPM-V only) const extracted = await extractInvoiceFromImages(images);
const extracted = await extractWithConsensus(images, testCase.name); console.log(` Extracted: ${extracted.invoice_number} | ${extracted.invoice_date} | ${extracted.total_amount} ${extracted.currency}`);
const endTime = Date.now(); const elapsedMs = Date.now() - startTime;
const elapsedMs = endTime - startTime;
processingTimes.push(elapsedMs); processingTimes.push(elapsedMs);
// Compare results
const result = compareInvoice(extracted, expected); const result = compareInvoice(extracted, expected);
if (result.match) { if (result.match) {
@@ -317,7 +474,6 @@ for (const testCase of testCases) {
result.errors.forEach((e) => console.log(` - ${e}`)); result.errors.forEach((e) => console.log(` - ${e}`));
} }
// Assert match
expect(result.match).toBeTrue(); expect(result.match).toBeTrue();
}); });
} }
@@ -326,18 +482,17 @@ tap.test('summary', async () => {
const totalInvoices = testCases.length; const totalInvoices = testCases.length;
const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0; const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
const totalTimeMs = processingTimes.reduce((a, b) => a + b, 0); const totalTimeMs = processingTimes.reduce((a, b) => a + b, 0);
const avgTimeMs = processingTimes.length > 0 ? totalTimeMs / processingTimes.length : 0; const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
const avgTimeSec = avgTimeMs / 1000;
const totalTimeSec = totalTimeMs / 1000;
console.log(`\n========================================`); console.log(`\n========================================`);
console.log(` Invoice Extraction Summary (MiniCPM)`); console.log(` Invoice Extraction Summary (MiniCPM)`);
console.log(`========================================`); console.log(`========================================`);
console.log(` Method: Multi-query (no_think)`);
console.log(` Passed: ${passedCount}/${totalInvoices}`); console.log(` Passed: ${passedCount}/${totalInvoices}`);
console.log(` Failed: ${failedCount}/${totalInvoices}`); console.log(` Failed: ${failedCount}/${totalInvoices}`);
console.log(` Accuracy: ${accuracy.toFixed(1)}%`); console.log(` Accuracy: ${accuracy.toFixed(1)}%`);
console.log(`----------------------------------------`); 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(` Avg per inv: ${avgTimeSec.toFixed(1)}s`);
console.log(`========================================\n`); console.log(`========================================\n`);
}); });

96
test/test.invoices.qwen3vl.ts
View File

@@ -1,10 +1,8 @@
/** /**
* Invoice extraction using Qwen3-VL 8B Vision (Direct) * Invoice extraction using Qwen3-VL 8B Vision (Direct)
* *
* Single-step pipeline: PDF → Images → Qwen3-VL → JSON * Multi-query approach: 5 parallel simple queries to avoid token exhaustion.
* Uses /no_think to disable reasoning mode for fast, direct responses. * Single pass, no consensus voting.
*
* Qwen3-VL outperforms PaddleOCR-VL on certain invoice formats.
*/ */
import { tap, expect } from '@git.zone/tstest/tapbundle'; import { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs'; import * as fs from 'fs';
@@ -67,11 +65,10 @@ async function queryField(images: string[], question: string): Promise<string> {
model: VISION_MODEL, model: VISION_MODEL,
messages: [{ messages: [{
role: 'user', role: 'user',
content: `/no_think\n${question} Reply with just the value, nothing else.`, content: `${question} Reply with just the value, nothing else.`,
images: images, images: images,
}], }],
stream: false, stream: false,
think: false,
options: { options: {
num_predict: 500, num_predict: 500,
temperature: 0.1, temperature: 0.1,
@@ -96,35 +93,80 @@ async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`); console.log(` [Vision] Processing ${images.length} page(s) with Qwen3-VL (multi-query)`);
// Query each field separately to avoid excessive thinking tokens // Query each field separately to avoid excessive thinking tokens
const [invoiceNum, invoiceDate, vendor, currency, amounts] = await Promise.all([ // Use explicit questions to avoid confusion between similar fields
queryField(images, 'What is the invoice number on this document?'), // Log each result as it comes in (not waiting for all to complete)
queryField(images, 'What is the invoice date? Format as YYYY-MM-DD.'), const queryAndLog = async (name: string, question: string): Promise<string> => {
queryField(images, 'What company issued this invoice?'), const result = await queryField(images, question);
queryField(images, 'What currency is used? Answer EUR, USD, or GBP.'), console.log(` [Query] ${name}: "${result}"`);
queryField(images, 'What are the net amount, VAT amount, and total amount? Format: net,vat,total'), 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'),
]); ]);
console.log(` [Vision] Got: ${invoiceNum} | ${invoiceDate} | ${vendor} | ${currency}`); // Parse amount from string (handles European format)
// Parse amounts (format: "net,vat,total" or similar)
const amountMatch = amounts.match(/([\d.,]+)/g) || [];
const parseAmount = (s: string): number => { const parseAmount = (s: string): number => {
if (!s) return 0; if (!s) return 0;
// Extract number from the response
const match = s.match(/([\d.,]+)/);
if (!match) return 0;
const numStr = match[1];
// Handle European format: 1.234,56 → 1234.56 // Handle European format: 1.234,56 → 1234.56
const normalized = s.includes(',') && s.indexOf(',') > s.lastIndexOf('.') const normalized = numStr.includes(',') && numStr.indexOf(',') > numStr.lastIndexOf('.')
? s.replace(/\./g, '').replace(',', '.') ? numStr.replace(/\./g, '').replace(',', '.')
: s.replace(/,/g, ''); : numStr.replace(/,/g, '');
return parseFloat(normalized) || 0; return parseFloat(normalized) || 0;
}; };
// Extract invoice number from potentially verbose response
const extractInvoiceNumber = (s: string): string => {
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
// Look for common invoice number patterns
const patterns = [
/\b([A-Z]{2,3}\d{10,})\b/i, // IEE2022006460244
/\b([A-Z]\d{8,})\b/i, // R0014359508
/\b(INV[-\s]?\d{4}[-\s]?\d+)\b/i, // INV-2024-001
/\b(\d{7,})\b/, // 1579087430
];
for (const pattern of patterns) {
const match = clean.match(pattern);
if (match) return match[1];
}
return clean.replace(/[^A-Z0-9-]/gi, '').trim() || clean;
};
// Extract date (YYYY-MM-DD) from response
const extractDate = (s: string): string => {
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
if (isoMatch) return isoMatch[1];
return clean.replace(/[^\d-]/g, '').trim();
};
// Extract currency
const extractCurrency = (s: string): string => {
const upper = s.toUpperCase();
if (upper.includes('EUR') || upper.includes('€')) return 'EUR';
if (upper.includes('USD') || upper.includes('$')) return 'USD';
if (upper.includes('GBP') || upper.includes('£')) return 'GBP';
return 'EUR';
};
return { return {
invoice_number: invoiceNum || '', invoice_number: extractInvoiceNumber(invoiceNum),
invoice_date: invoiceDate || '', invoice_date: extractDate(invoiceDate),
vendor_name: vendor || '', vendor_name: vendor.replace(/\*\*/g, '').replace(/`/g, '').trim() || '',
currency: (currency || 'EUR').toUpperCase().replace(/[^A-Z]/g, '').slice(0, 3) || 'EUR', currency: extractCurrency(currency),
net_amount: parseAmount(amountMatch[0] || ''), net_amount: parseAmount(netAmount),
vat_amount: parseAmount(amountMatch[1] || ''), vat_amount: parseAmount(vatAmount),
total_amount: parseAmount(amountMatch[2] || amountMatch[0] || ''), total_amount: parseAmount(totalAmount),
}; };
} }
@@ -296,7 +338,7 @@ tap.test('summary', async () => {
console.log(`\n======================================================`); console.log(`\n======================================================`);
console.log(` Invoice Extraction Summary (Qwen3-VL Vision)`); console.log(` Invoice Extraction Summary (Qwen3-VL Vision)`);
console.log(`======================================================`); console.log(`======================================================`);
console.log(` Method: Qwen3-VL 8B Direct Vision (/no_think)`); console.log(` Method: Multi-query (single pass)`);
console.log(` Passed: ${passedCount}/${total}`); console.log(` Passed: ${passedCount}/${total}`);
console.log(` Failed: ${failedCount}/${total}`); console.log(` Failed: ${failedCount}/${total}`);
console.log(` Accuracy: ${accuracy.toFixed(1)}%`); console.log(` Accuracy: ${accuracy.toFixed(1)}%`);