mongodb-memory-bank-mcp-v2
Version:
MongoDB-powered Memory Bank MCP server with hybrid search capabilities for AI assistants
354 lines (341 loc) • 14.2 kB
JavaScript
import { join } from 'path';
import { readFileSync, existsSync } from 'fs';
import { SearchToolSchema } from '../types/memory.js';
import { getMemoryCollection } from '../db/connection.js';
import { generateEmbedding } from '../embeddings/voyage.js';
import { logger } from '../utils/logger.js';
export async function searchTool(args) {
try {
const params = SearchToolSchema.parse(args);
const projectPath = params.projectPath || process.cwd();
// Read project config
const configPath = join(projectPath, '.memory-bank', 'config.json');
if (!existsSync(configPath)) {
return {
content: [
{
type: 'text',
text: 'Memory bank not initialized for this project. Run memory_bank/init first.',
},
],
};
}
const config = JSON.parse(readFileSync(configPath, 'utf-8'));
const collection = getMemoryCollection();
let results;
if (params.searchType === 'text') {
// Text-only search using Atlas Search
results = await collection
.aggregate([
{
$search: {
index: 'memory_text_index',
text: {
query: params.query,
path: ['content', 'fileName'],
},
},
},
{
$match: {
projectId: config.projectId,
},
},
{
$project: {
fileName: 1,
content: 1,
'metadata.type': 1,
score: { $meta: 'searchScore' },
},
},
{
$limit: params.limit,
},
])
.toArray();
}
else if (params.searchType === 'vector') {
// Vector-only search
const queryVector = await generateEmbedding(params.query);
// Check if vector index exists
const hasVectorData = await collection.findOne({
projectId: config.projectId,
contentVector: { $exists: true, $not: { $type: 'null' } },
});
if (!hasVectorData) {
return {
content: [
{
type: 'text',
text: 'No embeddings found. Run memory_bank/sync to generate embeddings first.',
},
],
};
}
// MongoDB Atlas Vector Search aggregation
// Use without filter and then match - works with basic vector index
results = await collection
.aggregate([
{
$vectorSearch: {
index: 'memory_vector_index',
path: 'contentVector',
queryVector: queryVector,
numCandidates: params.limit * 10,
limit: params.limit * 2,
},
},
{
$match: {
projectId: config.projectId,
},
},
{
$project: {
fileName: 1,
content: 1,
'metadata.type': 1,
score: { $meta: 'vectorSearchScore' },
},
},
{
$limit: params.limit,
},
])
.toArray();
}
else {
// Hybrid search using MongoDB's native $rankFusion
const queryVector = await generateEmbedding(params.query);
// Check if vector index exists
const hasVectorData = await collection.findOne({
projectId: config.projectId,
contentVector: { $exists: true, $not: { $type: 'null' } },
});
if (!hasVectorData) {
// Fall back to text search only
logger.warn('No embeddings found, falling back to text search');
params.searchType = 'text';
return searchTool(params);
}
// MongoDB native hybrid search with $rankFusion
results = await collection
.aggregate([
{
$rankFusion: {
input: {
pipelines: {
vectorPipeline: [
{
$vectorSearch: {
index: 'memory_vector_index',
path: 'contentVector',
queryVector: queryVector,
numCandidates: params.limit * 10,
limit: params.limit,
},
},
{
$match: {
projectId: config.projectId,
},
},
],
textPipeline: [
{
$search: {
index: 'memory_text_index',
text: {
query: params.query,
path: ['content', 'fileName'],
},
},
},
{
$match: {
projectId: config.projectId,
},
},
{
$limit: params.limit,
},
],
},
},
combination: {
weights: {
vectorPipeline: 0.7,
textPipeline: 0.3,
},
},
},
},
{
$addFields: {
score: { $meta: 'searchScore' }, // RRF score from $rankFusion
},
},
{
$project: {
fileName: 1,
content: 1,
'metadata.type': 1,
score: 1,
},
},
{
$limit: params.limit,
},
])
.toArray();
}
if (!results || results.length === 0) {
return {
content: [
{
type: 'text',
text: `🔍 No results found for query: "${params.query}"
💡 OPTIMIZATION TIPS:
1. 🔄 Run memory_bank/sync to generate/update embeddings
2. 🎯 Try different keywords or concepts
3. 🚀 Use hybrid search (default) for best results!
${params.searchType === 'hybrid' ?
`💎 MongoDB $rankFusion searched BOTH:
- Semantic meaning (what you're looking for conceptually)
- Exact keywords (what you typed)
No matches means this is truly NEW territory!` :
`🔧 You used ${params.searchType} search only. Try 'hybrid' for MongoDB's FULL POWER!`}`,
},
],
};
}
logger.info(`Search completed: ${results.length} results for query "${params.query}"`);
// Helper function to extract relevant context around search terms
const getSmartPreview = (content, query) => {
const lowerContent = content.toLowerCase();
const lowerQuery = query.toLowerCase();
const words = lowerQuery.split(/\s+/);
// Find the best matching section
let bestStart = 0;
let bestScore = 0;
// Check each position in the content
for (let i = 0; i < content.length - 150; i += 50) {
const section = lowerContent.substring(i, i + 300);
let score = 0;
// Count how many query words appear in this section
words.forEach(word => {
if (word && section.includes(word))
score++;
});
if (score > bestScore) {
bestScore = score;
bestStart = i;
}
}
// Extract the best section and clean it up
let preview = content.substring(bestStart, bestStart + 200);
// Try to start at sentence beginning
const sentenceStart = preview.indexOf('. ');
if (sentenceStart > 0 && sentenceStart < 50) {
preview = preview.substring(sentenceStart + 2);
}
// Clean up and add ellipsis if needed
preview = preview.replace(/\n+/g, ' ').trim();
if (bestStart > 0)
preview = '...' + preview;
if (bestStart + 200 < content.length)
preview = preview + '...';
return preview;
};
// Format results
const formattedResults = results.map((doc, index) => {
const preview = params.searchType === 'text' || params.searchType === 'hybrid'
? getSmartPreview(doc.content, params.query)
: doc.content.substring(0, 200).replace(/\n/g, ' ');
const scoreInfo = params.searchType === 'hybrid' && doc.vectorScore && doc.textScore
? ` (vector: ${doc.vectorScore.toFixed(3)}, text: ${doc.textScore.toFixed(3)})`
: '';
// Ensure score is always a number before calling toFixed
const scoreValue = typeof doc.score === 'number' ? doc.score : 0;
// Add content stats for AI context awareness
const contentStats = `[${(doc.content.length / 1024).toFixed(1)}KB, ${doc.metadata.version || 1} versions]`;
return `## ${index + 1}. ${doc.fileName} (${doc.metadata.type}) ${contentStats}
Score: ${scoreValue.toFixed(3)}${scoreInfo}
Preview: ${preview}`;
}).join('\n\n');
// Create search type explanation
let searchExplanation = '';
if (params.searchType === 'hybrid') {
searchExplanation = `💎 MongoDB $rankFusion Hybrid Search - The CROWN JEWEL!
🧪 Algorithm: 70% Semantic (Voyage AI) + 30% Keyword (Atlas Search)
🔄 Reciprocal Rank Fusion intelligently combined both approaches!
This search understood BOTH:
- What you meant conceptually (semantic vectors)
- The exact words you used (text matching)`;
}
else if (params.searchType === 'vector') {
searchExplanation = `🧠 Vector Search: Found semantically similar content using Voyage AI embeddings
💡 TIP: Use 'hybrid' search for even better results!`;
}
else {
searchExplanation = `📝 Text Search: Found keyword matches using Atlas Search
💡 TIP: Use 'hybrid' search to also find conceptually related content!`;
}
return {
content: [
{
type: 'text',
text: `🔍 MongoDB Search Results for: "${params.query}"
${searchExplanation}
📊 Found ${results.length} matches:
${formattedResults}
📍 NEXT STEPS:
1. 📖 Read the most relevant file: memory_bank/read --fileName "[filename]"
2. 🔄 Update your knowledge: memory_bank/update --fileName "[filename]"
3. 🚀 Create new features based on patterns found!
🏆 MongoDB ADVANTAGE: ONE database for:
- Operational data ✓
- Vector embeddings ✓
- Full-text search ✓
- Version history ✓
No external vector DB needed - MongoDB does it ALL!`,
},
],
};
}
catch (error) {
logger.error('Search tool error:', error);
// Provide user-friendly error messages
let errorMessage = 'Search failed: ';
const errorMsg = error instanceof Error ? error.message : String(error);
if (errorMsg.includes('no mongodb atlas search index found')) {
errorMessage = `Search indexes are being created. Please wait a moment and try again.
The following indexes are being set up:
- Vector search index for semantic search
- Atlas Search index for text search
This typically takes 1-2 minutes. Run memory_bank/sync to ensure indexes are created.`;
}
else if (errorMsg.includes('text index required')) {
errorMessage = `Text search is not available yet. The Atlas Search index is being created.
In the meantime, you can use:
- searchType: "vector" for semantic search
- Run memory_bank/sync to create search indexes`;
}
else if (error instanceof Error && 'code' in error && error.code === 6) {
errorMessage = 'Connection error. Please check your MongoDB connection and try again.';
}
else {
errorMessage += errorMsg || 'Unknown error occurred';
}
return {
content: [
{
type: 'text',
text: errorMessage,
},
],
};
}
}
//# sourceMappingURL=search.js.map