#!/usr/bin/env tsx
/**
* Diagnose why semantic search isn't working
*/
import { initDatabaseFromEnv } from './src/database/index.js';
import { MemoryCore } from './src/core/index.js';
import { EmbeddingService } from './src/utils/embeddings.js';
async function diagnoseSemanticSearch() {
console.log('🔍 Diagnosing Semantic Search Issues\n');
const db = initDatabaseFromEnv();
const apiKey = process.env.OPENAI_API_KEY;
const memoryCore = new MemoryCore(db, apiKey);
const embeddingService = new EmbeddingService(apiKey);
await memoryCore.initialize();
const userId = 'bob@matsuoka.com';
// 1. Check embedding statistics
console.log('1️⃣ EMBEDDING STATISTICS:');
const statsResult = await db.execute(`
SELECT
COUNT(*) as total,
COUNT(CASE WHEN embedding IS NOT NULL THEN 1 END) as with_any_embedding,
COUNT(CASE WHEN embedding IS NOT NULL AND json_array_length(embedding) > 0 THEN 1 END) as with_valid_embedding,
COUNT(CASE WHEN embedding = '[]' THEN 1 END) as with_empty_embedding
FROM memories
WHERE user_id = ?
`, [userId]);
const stats = statsResult.rows[0] as any;
console.log(` Total memories: ${stats.total}`);
console.log(` With embeddings: ${stats.with_valid_embedding} (${((stats.with_valid_embedding / stats.total) * 100).toFixed(1)}%)`);
console.log(` Empty embeddings: ${stats.with_empty_embedding}`);
// 2. Create test memories with embeddings
console.log('\n2️⃣ CREATING TEST MEMORIES:');
const testMemories = [
{ title: 'Machine Learning Basics', content: 'Neural networks, deep learning, and artificial intelligence fundamentals' },
{ title: 'Coffee Brewing Guide', content: 'How to make the perfect espresso with proper grinding and extraction' },
{ title: 'AI and Deep Learning', content: 'Advanced neural network architectures and transformer models' },
];
const createdIds = [];
for (const mem of testMemories) {
const result = await memoryCore.addMemory(
mem.title,
mem.content,
'semantic' as any,
{
userId,
generateEmbedding: true,
}
);
if (result.status === 'success' && result.data) {
const id = (result.data as any).id;
createdIds.push(id);
console.log(` ✅ Created: ${mem.title}`);
}
}
// Wait for embeddings
await new Promise(resolve => setTimeout(resolve, 2000));
// 3. Test semantic search
console.log('\n3️⃣ TESTING SEMANTIC SEARCH:');
const queries = [
'artificial intelligence and machine learning',
'neural networks',
'coffee espresso',
'transformer models'
];
for (const query of queries) {
console.log(`\n Query: "${query}"`);
// Try different strategies
const strategies = ['similarity', 'composite'];
for (const strategy of strategies) {
const searchResult = await memoryCore.searchMemories(query, {
userId,
strategy: strategy as any,
limit: 3,
threshold: 0.3,
});
if (searchResult.status === 'success' && searchResult.data) {
const results = searchResult.data as any[];
console.log(` Strategy ${strategy}: Found ${results.length} results`);
results.forEach((r, i) => {
console.log(` ${i + 1}. ${r.title}`);
});
} else {
console.log(` Strategy ${strategy}: Failed - ${searchResult.error}`);
}
}
}
// 4. Check if embeddings are being used
console.log('\n4️⃣ CHECKING EMBEDDING USAGE:');
// Generate embedding for test query
const testQuery = 'artificial intelligence';
const queryEmbedding = await embeddingService.generateEmbedding(testQuery);
console.log(` Query embedding generated: ${queryEmbedding.length} dimensions`);
// Get a memory with embedding
const memWithEmbedding = await db.execute(`
SELECT id, title, embedding
FROM memories
WHERE user_id = ?
AND embedding IS NOT NULL
AND json_array_length(embedding) > 0
LIMIT 1
`, [userId]);
if (memWithEmbedding.rows.length > 0) {
const row = memWithEmbedding.rows[0] as any;
const memEmbedding = JSON.parse(row.embedding);
// Calculate cosine similarity manually
const similarity = EmbeddingService.cosineSimilarity(queryEmbedding, memEmbedding);
console.log(` Manual similarity calculation: ${similarity.toFixed(4)}`);
console.log(` Memory: ${row.title}`);
}
// 5. Check metadata search
console.log('\n5️⃣ TESTING METADATA SEARCH:');
// Create memory with metadata
const metaResult = await memoryCore.addMemory(
'Test Metadata Memory',
'Testing metadata search functionality',
'semantic' as any,
{
userId,
metadata: {
projectId: 'alpha',
userId: 'user789',
category: 'test',
priority: 'high'
},
generateEmbedding: false,
}
);
if (metaResult.status === 'success' && metaResult.data) {
createdIds.push((metaResult.data as any).id);
// Test different metadata queries
const metaQueries = [
'metadata.projectId:alpha',
'metadata.userId:user789',
'projectId:alpha',
'userId:user789',
'category:test',
'metadata.category:test',
];
for (const query of metaQueries) {
const result = await memoryCore.searchMemories(query, { userId });
const found = result.status === 'success' && (result.data as any[]).length > 0;
console.log(` ${query}: ${found ? '✅ Found' : '❌ Not found'}`);
}
}
// Clean up
console.log('\n🧹 Cleaning up test data...');
for (const id of createdIds) {
await memoryCore.deleteMemory(id, { userId });
}
console.log('\n✅ Diagnosis complete!');
await db.disconnect();
process.exit(0);
}
diagnoseSemanticSearch().catch(console.error);
