feat(tests): revamp tests and remove legacy Dockerfiles: adopt JSON/consensus workflows, switch MiniCPM model, and delete deprecated Docker/test variants

This commit is contained in:
2026-01-18 13:56:46 +00:00
parent d6c97a9625
commit d91df70fff
15 changed files with 542 additions and 3142 deletions

View File

@@ -1,8 +1,10 @@
/**
* Invoice extraction test using MiniCPM-V only (visual extraction)
* Invoice extraction test using MiniCPM-V (visual extraction)
*
* Multi-query approach with thinking DISABLED for speed.
* Single pass, no consensus voting.
* Consensus approach:
* 1. Pass 1: Fast JSON extraction
* 2. Pass 2: Confirm with thinking enabled
* 3. If mismatch: repeat until consensus or max attempts
*/
import { tap, expect } from '@git.zone/tstest/tapbundle';
import * as fs from 'fs';
@@ -12,7 +14,7 @@ import * as os from 'os';
import { ensureMiniCpm } from './helpers/docker.js';
const OLLAMA_URL = 'http://localhost:11434';
const MODEL = 'minicpm-v:latest';
const MODEL = 'openbmb/minicpm-v4.5:q8_0';
interface IInvoice {
invoice_number: string;
@@ -33,7 +35,7 @@ function convertPdfToImages(pdfPath: string): string[] {
try {
execSync(
`convert -density 200 -quality 90 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
`convert -density 300 -quality 95 "${pdfPath}" -background white -alpha remove "${outputPattern}"`,
{ stdio: 'pipe' }
);
@@ -52,20 +54,35 @@ function convertPdfToImages(pdfPath: string): string[] {
}
}
const JSON_PROMPT = `Extract invoice data from this image. Return ONLY a JSON object with these exact fields:
{
"invoice_number": "the invoice number (not VAT ID, not customer ID)",
"invoice_date": "YYYY-MM-DD format",
"vendor_name": "company that issued the invoice",
"currency": "EUR, USD, or GBP",
"net_amount": 0.00,
"vat_amount": 0.00,
"total_amount": 0.00
}
Return only the JSON, no explanation.`;
/**
* Query MiniCPM-V for a single field (thinking disabled for speed)
* Query MiniCPM-V for JSON output (fast, no thinking)
*/
async function queryField(images: string[], question: string): Promise<string> {
const response = await fetch(`${OLLAMA_URL}/api/generate`, {
async function queryJsonFast(images: string[]): Promise<string> {
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
model: MODEL,
prompt: `/no_think\n${question}`,
images: images,
messages: [{
role: 'user',
content: JSON_PROMPT,
images: images,
}],
stream: false,
options: {
num_predict: 500,
num_predict: 1000,
temperature: 0.1,
},
}),
@@ -76,263 +93,223 @@ async function queryField(images: string[], question: string): Promise<string> {
}
const data = await response.json();
const content = (data.response || '').trim();
// Return full content (no thinking to filter)
return content;
return (data.message?.content || '').trim();
}
/**
* Extract invoice data using multiple queries with validation
* Query MiniCPM-V for JSON output with thinking enabled (slower, more accurate)
*/
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
console.log(` [Vision] Processing ${images.length} page(s) with MiniCPM-V (multi-query + validation)`);
async function queryJsonWithThinking(images: string[]): Promise<string> {
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
model: MODEL,
messages: [{
role: 'user',
content: `Think carefully about this invoice image, then ${JSON_PROMPT}`,
images: images,
}],
stream: false,
options: {
num_predict: 2000,
temperature: 0.1,
},
}),
});
// 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;
};
if (!response.ok) {
throw new Error(`Ollama API error: ${response.status}`);
}
// 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<<<`),
]);
const data = await response.json();
return (data.message?.content || '').trim();
}
// 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<<<`),
/**
* Parse amount from string (handles European format)
*/
function parseAmount(s: string | number | undefined): number {
if (s === undefined || s === null) return 0;
if (typeof s === 'number') return s;
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;
}
// 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<<<`),
/**
* Extract invoice number from potentially verbose response
*/
function extractInvoiceNumber(s: string | undefined): string {
if (!s) return '';
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
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;
}
queryAndLog('Vendor ', `What company ISSUED this invoice (seller, not buyer)?
Look at letterhead/logo at top.
Return ONLY the company name: >>>Adobe Inc.<<<`),
/**
* Extract date (YYYY-MM-DD) from response
*/
function extractDate(s: string | undefined): string {
if (!s) return '';
let clean = s.replace(/\*\*/g, '').replace(/`/g, '').trim();
const isoMatch = clean.match(/(\d{4}-\d{2}-\d{2})/);
if (isoMatch) return isoMatch[1];
// Try DD/MM/YYYY or DD.MM.YYYY
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 clean.replace(/[^\d-]/g, '').trim();
}
queryAndLog('Currency ', `What currency symbol appears next to amounts? € $ or £?
Return the 3-letter code: >>>EUR<<<`),
/**
* Extract currency
*/
function extractCurrency(s: string | undefined): string {
if (!s) return 'EUR';
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';
}
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<<<`),
/**
* Extract JSON from response (handles markdown code blocks)
*/
function extractJsonFromResponse(response: string): Record<string, unknown> | null {
// Try to find JSON in markdown code block
const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
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;
try {
return JSON.parse(jsonStr);
} catch {
// Try to find JSON object pattern
const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
if (jsonMatch) {
try {
return JSON.parse(jsonMatch[0]);
} catch {
return null;
}
}
// 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}`);
}
}
/**
* Parse JSON response into IInvoice
*/
function parseJsonToInvoice(response: string): IInvoice | null {
const parsed = extractJsonFromResponse(response);
if (!parsed) return null;
return {
invoice_number: invoiceNumber,
invoice_date: finalDate,
vendor_name: extractVendor(vendor),
currency: extractCurrency(currency),
net_amount: net,
vat_amount: vat,
total_amount: total,
invoice_number: extractInvoiceNumber(String(parsed.invoice_number || '')),
invoice_date: extractDate(String(parsed.invoice_date || '')),
vendor_name: String(parsed.vendor_name || '').replace(/\*\*/g, '').replace(/`/g, '').trim(),
currency: extractCurrency(String(parsed.currency || '')),
net_amount: parseAmount(parsed.net_amount as string | number),
vat_amount: parseAmount(parsed.vat_amount as string | number),
total_amount: parseAmount(parsed.total_amount as string | number),
};
}
/**
* Compare two invoices for consensus (key fields must match)
*/
function invoicesMatch(a: IInvoice, b: IInvoice): boolean {
const numMatch = a.invoice_number.toLowerCase() === b.invoice_number.toLowerCase();
const dateMatch = a.invoice_date === b.invoice_date;
const totalMatch = Math.abs(a.total_amount - b.total_amount) < 0.02;
return numMatch && dateMatch && totalMatch;
}
/**
* Extract invoice data using consensus approach:
* 1. Pass 1: Fast JSON extraction
* 2. Pass 2: Confirm with thinking enabled
* 3. If mismatch: repeat until consensus or max 5 attempts
*/
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
console.log(` [Vision] Processing ${images.length} page(s) with ${MODEL} (consensus)`);
const MAX_ATTEMPTS = 5;
let attempt = 0;
while (attempt < MAX_ATTEMPTS) {
attempt++;
console.log(` [Attempt ${attempt}/${MAX_ATTEMPTS}]`);
// PASS 1: Fast JSON extraction
console.log(` [Pass 1] Fast extraction...`);
const fastResponse = await queryJsonFast(images);
const fastInvoice = parseJsonToInvoice(fastResponse);
if (!fastInvoice) {
console.log(` [Pass 1] JSON parsing failed, retrying...`);
continue;
}
console.log(` [Pass 1] Result: ${fastInvoice.invoice_number} | ${fastInvoice.invoice_date} | ${fastInvoice.total_amount} ${fastInvoice.currency}`);
// PASS 2: Confirm with thinking
console.log(` [Pass 2] Thinking confirmation...`);
const thinkResponse = await queryJsonWithThinking(images);
const thinkInvoice = parseJsonToInvoice(thinkResponse);
if (!thinkInvoice) {
console.log(` [Pass 2] JSON parsing failed, retrying...`);
continue;
}
console.log(` [Pass 2] Result: ${thinkInvoice.invoice_number} | ${thinkInvoice.invoice_date} | ${thinkInvoice.total_amount} ${thinkInvoice.currency}`);
// Check consensus
if (invoicesMatch(fastInvoice, thinkInvoice)) {
console.log(` [Consensus] MATCH - using result`);
return thinkInvoice; // Prefer thinking result
}
console.log(` [Consensus] MISMATCH - repeating...`);
console.log(` Fast: ${fastInvoice.invoice_number} | ${fastInvoice.invoice_date} | ${fastInvoice.total_amount}`);
console.log(` Think: ${thinkInvoice.invoice_number} | ${thinkInvoice.invoice_date} | ${thinkInvoice.total_amount}`);
}
// Max attempts reached - do one final thinking pass and use that
console.log(` [Final] Max attempts reached, using final thinking pass`);
const finalResponse = await queryJsonWithThinking(images);
const finalInvoice = parseJsonToInvoice(finalResponse);
if (finalInvoice) {
console.log(` [Final] Result: ${finalInvoice.invoice_number} | ${finalInvoice.invoice_date} | ${finalInvoice.total_amount} ${finalInvoice.currency}`);
return finalInvoice;
}
// Return empty invoice if all else fails
console.log(` [Final] All parsing failed, returning empty`);
return {
invoice_number: '',
invoice_date: '',
vendor_name: '',
currency: 'EUR',
net_amount: 0,
vat_amount: 0,
total_amount: 0,
};
}
@@ -485,9 +462,9 @@ tap.test('summary', async () => {
const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
console.log(`\n========================================`);
console.log(` Invoice Extraction Summary (MiniCPM)`);
console.log(` Invoice Extraction Summary (${MODEL})`);
console.log(`========================================`);
console.log(` Method: Multi-query (no_think)`);
console.log(` Method: Consensus (fast + thinking)`);
console.log(` Passed: ${passedCount}/${totalInvoices}`);
console.log(` Failed: ${failedCount}/${totalInvoices}`);
console.log(` Accuracy: ${accuracy.toFixed(1)}%`);