feat(tests): integrate SmartAi/DualAgentOrchestrator into extraction tests and add JSON self-validation
This commit is contained in:
@@ -1,10 +1,10 @@
|
||||
/**
|
||||
* Invoice extraction test using MiniCPM-V (visual extraction)
|
||||
* Invoice extraction test using MiniCPM-V via smartagent DualAgentOrchestrator
|
||||
*
|
||||
* Consensus approach:
|
||||
* 1. Pass 1: Fast JSON extraction
|
||||
* 2. Pass 2: Confirm with thinking enabled
|
||||
* 3. If mismatch: repeat until consensus or max attempts
|
||||
* Uses vision-capable orchestrator with JsonValidatorTool for self-validation:
|
||||
* 1. Pass images to the orchestrator
|
||||
* 2. Driver extracts invoice data and validates JSON before completing
|
||||
* 3. If validation fails, driver retries within the same task
|
||||
*/
|
||||
import { tap, expect } from '@git.zone/tstest/tapbundle';
|
||||
import * as fs from 'fs';
|
||||
@@ -12,6 +12,8 @@ import * as path from 'path';
|
||||
import { execSync } from 'child_process';
|
||||
import * as os from 'os';
|
||||
import { ensureMiniCpm } from './helpers/docker.js';
|
||||
import { SmartAi } from '@push.rocks/smartai';
|
||||
import { DualAgentOrchestrator, JsonValidatorTool } from '@push.rocks/smartagent';
|
||||
|
||||
const OLLAMA_URL = 'http://localhost:11434';
|
||||
const MODEL = 'openbmb/minicpm-v4.5:q8_0';
|
||||
@@ -26,6 +28,10 @@ interface IInvoice {
|
||||
total_amount: number;
|
||||
}
|
||||
|
||||
// SmartAi instance and orchestrator (initialized in setup)
|
||||
let smartAi: SmartAi;
|
||||
let orchestrator: DualAgentOrchestrator;
|
||||
|
||||
/**
|
||||
* Convert PDF to PNG images using ImageMagick
|
||||
*/
|
||||
@@ -54,7 +60,9 @@ function convertPdfToImages(pdfPath: string): string[] {
|
||||
}
|
||||
}
|
||||
|
||||
const JSON_PROMPT = `Extract invoice data from this image. Return ONLY a JSON object with these exact fields:
|
||||
const EXTRACTION_PROMPT = `Extract invoice data from the provided image(s).
|
||||
|
||||
IMPORTANT: You must output a valid JSON object with these exact fields:
|
||||
{
|
||||
"invoice_number": "the invoice number (not VAT ID, not customer ID)",
|
||||
"invoice_date": "YYYY-MM-DD format",
|
||||
@@ -64,150 +72,16 @@ const JSON_PROMPT = `Extract invoice data from this image. Return ONLY a JSON ob
|
||||
"vat_amount": 0.00,
|
||||
"total_amount": 0.00
|
||||
}
|
||||
Return only the JSON, no explanation.`;
|
||||
|
||||
/**
|
||||
* Query MiniCPM-V for JSON output (fast, no thinking) with streaming
|
||||
*/
|
||||
async function queryJsonFast(images: string[]): Promise<string> {
|
||||
const startTime = Date.now();
|
||||
process.stdout.write(` [Fast] `);
|
||||
Before completing, use the json.validate tool to verify your output is valid JSON with all required fields.
|
||||
|
||||
const response = await fetch(`${OLLAMA_URL}/api/chat`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({
|
||||
model: MODEL,
|
||||
messages: [{
|
||||
role: 'user',
|
||||
content: JSON_PROMPT,
|
||||
images: images,
|
||||
}],
|
||||
stream: true,
|
||||
options: {
|
||||
num_ctx: 32768,
|
||||
num_predict: 1000,
|
||||
temperature: 0.1,
|
||||
},
|
||||
}),
|
||||
});
|
||||
<tool_call>
|
||||
<tool>json</tool>
|
||||
<action>validate</action>
|
||||
<params>{"jsonString": "YOUR_JSON_HERE", "requiredFields": ["invoice_number", "invoice_date", "vendor_name", "currency", "net_amount", "vat_amount", "total_amount"]}</params>
|
||||
</tool_call>
|
||||
|
||||
if (!response.ok) {
|
||||
throw new Error(`Ollama API error: ${response.status}`);
|
||||
}
|
||||
|
||||
let content = '';
|
||||
const reader = response.body!.getReader();
|
||||
const decoder = new TextDecoder();
|
||||
|
||||
try {
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
for (const line of chunk.split('\n').filter(l => l.trim())) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
const token = json.message?.content || '';
|
||||
if (token) {
|
||||
process.stdout.write(token);
|
||||
content += token;
|
||||
}
|
||||
} catch {
|
||||
// Ignore parse errors for partial chunks
|
||||
}
|
||||
}
|
||||
}
|
||||
} finally {
|
||||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||||
process.stdout.write(` (${elapsed}s)\n`);
|
||||
}
|
||||
|
||||
return content.trim();
|
||||
}
|
||||
|
||||
/**
|
||||
* Query MiniCPM-V for JSON output with thinking enabled (slower, more accurate) with streaming
|
||||
*/
|
||||
async function queryJsonWithThinking(images: string[]): Promise<string> {
|
||||
const startTime = Date.now();
|
||||
process.stdout.write(` [Think] `);
|
||||
|
||||
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: true,
|
||||
options: {
|
||||
num_ctx: 32768,
|
||||
num_predict: 2000,
|
||||
temperature: 0.1,
|
||||
},
|
||||
}),
|
||||
});
|
||||
|
||||
if (!response.ok) {
|
||||
throw new Error(`Ollama API error: ${response.status}`);
|
||||
}
|
||||
|
||||
let content = '';
|
||||
let thinkingContent = '';
|
||||
let thinkingStarted = false;
|
||||
let outputStarted = false;
|
||||
const reader = response.body!.getReader();
|
||||
const decoder = new TextDecoder();
|
||||
|
||||
try {
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
const chunk = decoder.decode(value, { stream: true });
|
||||
for (const line of chunk.split('\n').filter(l => l.trim())) {
|
||||
try {
|
||||
const json = JSON.parse(line);
|
||||
|
||||
// Stream thinking tokens
|
||||
const thinking = json.message?.thinking || '';
|
||||
if (thinking) {
|
||||
if (!thinkingStarted) {
|
||||
process.stdout.write(`THINKING: `);
|
||||
thinkingStarted = true;
|
||||
}
|
||||
process.stdout.write(thinking);
|
||||
thinkingContent += thinking;
|
||||
}
|
||||
|
||||
// Stream content tokens
|
||||
const token = json.message?.content || '';
|
||||
if (token) {
|
||||
if (!outputStarted) {
|
||||
if (thinkingStarted) process.stdout.write('\n [Think] ');
|
||||
process.stdout.write(`OUTPUT: `);
|
||||
outputStarted = true;
|
||||
}
|
||||
process.stdout.write(token);
|
||||
content += token;
|
||||
}
|
||||
} catch {
|
||||
// Ignore parse errors for partial chunks
|
||||
}
|
||||
}
|
||||
}
|
||||
} finally {
|
||||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||||
process.stdout.write(` (${elapsed}s)\n`);
|
||||
}
|
||||
|
||||
return content.trim();
|
||||
}
|
||||
Only complete the task after validation passes. Output the final JSON in <task_complete> tags.`;
|
||||
|
||||
/**
|
||||
* Parse amount from string (handles European format)
|
||||
@@ -273,9 +147,31 @@ function extractCurrency(s: string | undefined): string {
|
||||
}
|
||||
|
||||
/**
|
||||
* Extract JSON from response (handles markdown code blocks)
|
||||
* Extract JSON from response (handles markdown code blocks and task_complete tags)
|
||||
*/
|
||||
function extractJsonFromResponse(response: string): Record<string, unknown> | null {
|
||||
// Try to find JSON in task_complete tags
|
||||
const completeMatch = response.match(/<task_complete>([\s\S]*?)<\/task_complete>/);
|
||||
if (completeMatch) {
|
||||
const content = completeMatch[1].trim();
|
||||
// Try to find JSON in the content
|
||||
const codeBlockMatch = content.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||||
const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : content;
|
||||
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 JSON in markdown code block
|
||||
const codeBlockMatch = response.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||||
const jsonStr = codeBlockMatch ? codeBlockMatch[1].trim() : response.trim();
|
||||
@@ -315,76 +211,27 @@ function parseJsonToInvoice(response: string): IInvoice | null {
|
||||
}
|
||||
|
||||
/**
|
||||
* 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
|
||||
* Extract invoice data using smartagent orchestrator with vision
|
||||
*/
|
||||
async function extractInvoiceFromImages(images: string[]): Promise<IInvoice> {
|
||||
console.log(` [Vision] Processing ${images.length} page(s) with ${MODEL} (consensus)`);
|
||||
console.log(` [Vision] Processing ${images.length} page(s) with smartagent DualAgentOrchestrator`);
|
||||
|
||||
const MAX_ATTEMPTS = 5;
|
||||
let attempt = 0;
|
||||
const startTime = Date.now();
|
||||
|
||||
while (attempt < MAX_ATTEMPTS) {
|
||||
attempt++;
|
||||
console.log(` [Attempt ${attempt}/${MAX_ATTEMPTS}]`);
|
||||
const result = await orchestrator.run(EXTRACTION_PROMPT, { images });
|
||||
|
||||
// PASS 1: Fast JSON extraction
|
||||
console.log(` [Pass 1] Fast extraction...`);
|
||||
const fastResponse = await queryJsonFast(images);
|
||||
const fastInvoice = parseJsonToInvoice(fastResponse);
|
||||
const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
|
||||
console.log(` [Vision] Completed in ${elapsed}s (${result.iterations} iterations, status: ${result.status})`);
|
||||
|
||||
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}`);
|
||||
const invoice = parseJsonToInvoice(result.result);
|
||||
|
||||
// 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}`);
|
||||
if (invoice) {
|
||||
console.log(` [Result] ${invoice.invoice_number} | ${invoice.invoice_date} | ${invoice.total_amount} ${invoice.currency}`);
|
||||
return invoice;
|
||||
}
|
||||
|
||||
// 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 empty invoice if parsing failed
|
||||
console.log(` [Result] Parsing failed, returning empty invoice`);
|
||||
return {
|
||||
invoice_number: '',
|
||||
invoice_date: '',
|
||||
@@ -493,6 +340,79 @@ tap.test('setup: ensure Docker containers are running', async () => {
|
||||
console.log('\n[Setup] All containers ready!\n');
|
||||
});
|
||||
|
||||
tap.test('setup: initialize smartagent orchestrator', async () => {
|
||||
console.log('[Setup] Initializing SmartAi and DualAgentOrchestrator...');
|
||||
|
||||
smartAi = new SmartAi({
|
||||
ollama: {
|
||||
baseUrl: OLLAMA_URL,
|
||||
model: MODEL,
|
||||
defaultOptions: {
|
||||
num_ctx: 32768,
|
||||
temperature: 0.1,
|
||||
},
|
||||
defaultTimeout: 300000, // 5 minutes for vision tasks
|
||||
},
|
||||
});
|
||||
|
||||
await smartAi.start();
|
||||
|
||||
orchestrator = new DualAgentOrchestrator({
|
||||
smartAiInstance: smartAi,
|
||||
defaultProvider: 'ollama',
|
||||
guardianPolicyPrompt: `You are a Guardian agent overseeing invoice extraction tasks.
|
||||
|
||||
APPROVE all tool calls that:
|
||||
- Use the json.validate action to verify JSON output
|
||||
- Are reasonable attempts to complete the extraction task
|
||||
|
||||
REJECT tool calls that:
|
||||
- Attempt to access external resources
|
||||
- Try to execute arbitrary code
|
||||
- Are clearly unrelated to invoice extraction`,
|
||||
driverSystemMessage: `You are an AI assistant that extracts invoice data from images.
|
||||
|
||||
Your task is to analyze invoice images and extract structured data.
|
||||
You have access to a json.validate tool to verify your JSON output.
|
||||
|
||||
IMPORTANT: Always validate your JSON before completing the task.
|
||||
|
||||
## Tool Usage Format
|
||||
When you need to validate JSON, output:
|
||||
|
||||
<tool_call>
|
||||
<tool>json</tool>
|
||||
<action>validate</action>
|
||||
<params>{"jsonString": "YOUR_JSON", "requiredFields": ["invoice_number", "invoice_date", "vendor_name", "currency", "net_amount", "vat_amount", "total_amount"]}</params>
|
||||
</tool_call>
|
||||
|
||||
## Completion Format
|
||||
After validation passes, complete the task:
|
||||
|
||||
<task_complete>
|
||||
{"invoice_number": "...", "invoice_date": "YYYY-MM-DD", ...}
|
||||
</task_complete>`,
|
||||
maxIterations: 5,
|
||||
maxConsecutiveRejections: 3,
|
||||
onToken: (token, source) => {
|
||||
if (source === 'driver') {
|
||||
process.stdout.write(token);
|
||||
}
|
||||
},
|
||||
onProgress: (event) => {
|
||||
if (event.logLevel === 'error') {
|
||||
console.error(event.logMessage);
|
||||
}
|
||||
},
|
||||
});
|
||||
|
||||
// Register the JsonValidatorTool
|
||||
orchestrator.registerTool(new JsonValidatorTool());
|
||||
|
||||
await orchestrator.start();
|
||||
console.log('[Setup] Orchestrator initialized!\n');
|
||||
});
|
||||
|
||||
tap.test('should have MiniCPM-V model loaded', async () => {
|
||||
const response = await fetch(`${OLLAMA_URL}/api/tags`);
|
||||
const data = await response.json();
|
||||
@@ -501,7 +421,7 @@ tap.test('should have MiniCPM-V model loaded', async () => {
|
||||
});
|
||||
|
||||
const testCases = findTestCases();
|
||||
console.log(`\nFound ${testCases.length} invoice test cases (MiniCPM-V)\n`);
|
||||
console.log(`\nFound ${testCases.length} invoice test cases (smartagent + MiniCPM-V)\n`);
|
||||
|
||||
let passedCount = 0;
|
||||
let failedCount = 0;
|
||||
@@ -538,6 +458,13 @@ for (const testCase of testCases) {
|
||||
});
|
||||
}
|
||||
|
||||
tap.test('cleanup: stop orchestrator', async () => {
|
||||
if (orchestrator) {
|
||||
await orchestrator.stop();
|
||||
}
|
||||
console.log('[Cleanup] Orchestrator stopped');
|
||||
});
|
||||
|
||||
tap.test('summary', async () => {
|
||||
const totalInvoices = testCases.length;
|
||||
const accuracy = totalInvoices > 0 ? (passedCount / totalInvoices) * 100 : 0;
|
||||
@@ -545,9 +472,10 @@ tap.test('summary', async () => {
|
||||
const avgTimeSec = processingTimes.length > 0 ? totalTimeMs / processingTimes.length / 1000 : 0;
|
||||
|
||||
console.log(`\n========================================`);
|
||||
console.log(` Invoice Extraction Summary (${MODEL})`);
|
||||
console.log(` Invoice Extraction Summary`);
|
||||
console.log(` (smartagent + ${MODEL})`);
|
||||
console.log(`========================================`);
|
||||
console.log(` Method: Consensus (fast + thinking)`);
|
||||
console.log(` Method: DualAgentOrchestrator with vision`);
|
||||
console.log(` Passed: ${passedCount}/${totalInvoices}`);
|
||||
console.log(` Failed: ${failedCount}/${totalInvoices}`);
|
||||
console.log(` Accuracy: ${accuracy.toFixed(1)}%`);
|
||||
|
||||
Reference in New Issue
Block a user