Implement hybrid search: Chroma semantic + SQLite temporal
Core implementation: - Added Chroma MCP client integration to search-server.ts - Implemented queryChroma() helper with Python dict parsing - Added VECTOR_DB_DIR constant to paths.ts - Added SessionStore.getObservationsByIds() method Search handlers updated: - search_observations: Semantic-first with 90-day temporal filter - find_by_concept/type/file: Metadata-first, semantic-enhanced ranking - All handlers fall back to FTS5 if Chroma unavailable Technical details: - Direct MCP client usage (no abstractions) - Regex parsing of Chroma Python dict responses - Semantic ranking preserved in final results - Graceful degradation to FTS5-only search 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
Alex Newman
+305
-14
@@ -5,6 +5,8 @@
|
||||
|
||||
import { Server } from '@modelcontextprotocol/sdk/server/index.js';
|
||||
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
|
||||
import { Client } from '@modelcontextprotocol/sdk/client/index.js';
|
||||
import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';
|
||||
import {
|
||||
CallToolRequestSchema,
|
||||
ListToolsRequestSchema,
|
||||
@@ -15,10 +17,14 @@ import { basename } from 'path';
|
||||
import { SessionSearch } from '../services/sqlite/SessionSearch.js';
|
||||
import { SessionStore } from '../services/sqlite/SessionStore.js';
|
||||
import { ObservationSearchResult, SessionSummarySearchResult, UserPromptSearchResult } from '../services/sqlite/types.js';
|
||||
import { VECTOR_DB_DIR } from '../shared/paths.js';
|
||||
|
||||
// Initialize search instance
|
||||
// Initialize search instances
|
||||
let search: SessionSearch;
|
||||
let store: SessionStore;
|
||||
let chromaClient: Client | null = null;
|
||||
const COLLECTION_NAME = 'cm__claude-mem';
|
||||
|
||||
try {
|
||||
search = new SessionSearch();
|
||||
store = new SessionStore();
|
||||
@@ -27,6 +33,95 @@ try {
|
||||
process.exit(1);
|
||||
}
|
||||
|
||||
/**
|
||||
* Query Chroma vector database via MCP
|
||||
* Parses Python dict-like responses from Chroma MCP server
|
||||
*/
|
||||
async function queryChroma(
|
||||
query: string,
|
||||
limit: number,
|
||||
whereFilter?: Record<string, any>
|
||||
): Promise<{ ids: number[]; distances: number[]; metadatas: any[] }> {
|
||||
if (!chromaClient) {
|
||||
throw new Error('Chroma client not initialized');
|
||||
}
|
||||
|
||||
const result = await chromaClient.callTool({
|
||||
name: 'chroma_query_documents',
|
||||
arguments: {
|
||||
collection_name: COLLECTION_NAME,
|
||||
query_texts: [query],
|
||||
n_results: limit,
|
||||
include: ['documents', 'metadatas', 'distances'],
|
||||
where: whereFilter
|
||||
}
|
||||
});
|
||||
|
||||
const resultText = result.content[0]?.text || '';
|
||||
|
||||
// Parse Python dict-like output using regex
|
||||
// Format: {'ids': [[...]], 'distances': [[...]], 'metadatas': [[...]]}
|
||||
|
||||
// Extract IDs (nested array format)
|
||||
const idsMatch = resultText.match(/'ids':\s*\[\[(.*?)\]\]/s);
|
||||
const ids: number[] = [];
|
||||
if (idsMatch) {
|
||||
const idsContent = idsMatch[1];
|
||||
// Match quoted strings (Chroma doc IDs like 'obs_123_title')
|
||||
const idMatches = idsContent.match(/'([^']*(?:\\'[^']*)*)'/g) || [];
|
||||
for (const idMatch of idMatches) {
|
||||
const docId = idMatch.slice(1, -1);
|
||||
// Extract sqlite_id from document ID (format: obs_{id}_title)
|
||||
const sqliteIdMatch = docId.match(/obs_(\d+)_/);
|
||||
if (sqliteIdMatch) {
|
||||
const sqliteId = parseInt(sqliteIdMatch[1], 10);
|
||||
if (!ids.includes(sqliteId)) {
|
||||
ids.push(sqliteId);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Extract distances (nested array format)
|
||||
const distancesMatch = resultText.match(/'distances':\s*\[\[([\d.,\s]+)\]\]/s);
|
||||
const distances: number[] = [];
|
||||
if (distancesMatch) {
|
||||
const distancesContent = distancesMatch[1];
|
||||
const distanceValues = distancesContent.split(',').map(d => parseFloat(d.trim())).filter(d => !isNaN(d));
|
||||
distances.push(...distanceValues);
|
||||
}
|
||||
|
||||
// Extract metadatas (nested array format)
|
||||
const metasMatch = resultText.match(/'metadatas':\s*\[\[(.*?)\]\]/s);
|
||||
const metadatas: any[] = [];
|
||||
if (metasMatch) {
|
||||
const metasContent = metasMatch[1];
|
||||
// Parse each metadata dict
|
||||
const metaObjMatches = metasContent.match(/\{[^}]+\}/g) || [];
|
||||
for (const metaStr of metaObjMatches) {
|
||||
const meta: any = {};
|
||||
// Extract sqlite_id
|
||||
const sqliteIdMatch = metaStr.match(/'sqlite_id':\s*(\d+)/);
|
||||
if (sqliteIdMatch) {
|
||||
meta.sqlite_id = parseInt(sqliteIdMatch[1], 10);
|
||||
}
|
||||
// Extract type
|
||||
const typeMatch = metaStr.match(/'type':\s*'([^']+)'/);
|
||||
if (typeMatch) {
|
||||
meta.type = typeMatch[1];
|
||||
}
|
||||
// Extract created_at_epoch
|
||||
const epochMatch = metaStr.match(/'created_at_epoch':\s*(\d+)/);
|
||||
if (epochMatch) {
|
||||
meta.created_at_epoch = parseInt(epochMatch[1], 10);
|
||||
}
|
||||
metadatas.push(meta);
|
||||
}
|
||||
}
|
||||
|
||||
return { ids, distances, metadatas };
|
||||
}
|
||||
|
||||
/**
|
||||
* Format search tips footer
|
||||
*/
|
||||
@@ -286,7 +381,45 @@ const tools = [
|
||||
handler: async (args: any) => {
|
||||
try {
|
||||
const { query, format = 'index', ...options } = args;
|
||||
const results = search.searchObservations(query, options);
|
||||
let results: ObservationSearchResult[] = [];
|
||||
|
||||
// Hybrid search: Try Chroma semantic search first, fall back to FTS5
|
||||
if (chromaClient) {
|
||||
try {
|
||||
console.error('[search-server] Using hybrid semantic search (Chroma + SQLite)');
|
||||
|
||||
// Step 1: Chroma semantic search (top 100)
|
||||
const chromaResults = await queryChroma(query, 100);
|
||||
console.error(`[search-server] Chroma returned ${chromaResults.ids.length} semantic matches`);
|
||||
|
||||
if (chromaResults.ids.length > 0) {
|
||||
// Step 2: Filter by recency (90 days)
|
||||
const ninetyDaysAgo = Math.floor(Date.now() / 1000) - (90 * 24 * 60 * 60);
|
||||
const recentIds = chromaResults.ids.filter((id, idx) => {
|
||||
const meta = chromaResults.metadatas[idx];
|
||||
return meta && meta.created_at_epoch > ninetyDaysAgo;
|
||||
});
|
||||
|
||||
console.error(`[search-server] ${recentIds.length} results within 90-day window`);
|
||||
|
||||
// Step 3: Hydrate from SQLite in temporal order
|
||||
if (recentIds.length > 0) {
|
||||
const limit = options.limit || 20;
|
||||
results = store.getObservationsByIds(recentIds, { orderBy: 'date_desc', limit });
|
||||
console.error(`[search-server] Hydrated ${results.length} observations from SQLite`);
|
||||
}
|
||||
}
|
||||
} catch (chromaError: any) {
|
||||
console.error('[search-server] Chroma query failed, falling back to FTS5:', chromaError.message);
|
||||
// Fall through to FTS5 fallback
|
||||
}
|
||||
}
|
||||
|
||||
// Fall back to FTS5 if Chroma unavailable or returned no results
|
||||
if (results.length === 0) {
|
||||
console.error('[search-server] Using FTS5 keyword search');
|
||||
results = search.searchObservations(query, options);
|
||||
}
|
||||
|
||||
if (results.length === 0) {
|
||||
return {
|
||||
@@ -400,7 +533,50 @@ const tools = [
|
||||
handler: async (args: any) => {
|
||||
try {
|
||||
const { concept, format = 'index', ...filters } = args;
|
||||
const results = search.findByConcept(concept, filters);
|
||||
let results: ObservationSearchResult[] = [];
|
||||
|
||||
// Metadata-first, semantic-enhanced search
|
||||
if (chromaClient) {
|
||||
try {
|
||||
console.error('[search-server] Using metadata-first + semantic ranking for concept search');
|
||||
|
||||
// Step 1: SQLite metadata filter (get all IDs with this concept)
|
||||
const metadataResults = search.findByConcept(concept, filters);
|
||||
console.error(`[search-server] Found ${metadataResults.length} observations with concept "${concept}"`);
|
||||
|
||||
if (metadataResults.length > 0) {
|
||||
// Step 2: Chroma semantic ranking (rank by relevance to concept)
|
||||
const ids = metadataResults.map(obs => obs.id);
|
||||
const chromaResults = await queryChroma(concept, Math.min(ids.length, 100));
|
||||
|
||||
// Intersect: Keep only IDs that passed metadata filter, in semantic rank order
|
||||
const rankedIds: number[] = [];
|
||||
for (const chromaId of chromaResults.ids) {
|
||||
if (ids.includes(chromaId) && !rankedIds.includes(chromaId)) {
|
||||
rankedIds.push(chromaId);
|
||||
}
|
||||
}
|
||||
|
||||
console.error(`[search-server] Chroma ranked ${rankedIds.length} results by semantic relevance`);
|
||||
|
||||
// Step 3: Hydrate in semantic rank order
|
||||
if (rankedIds.length > 0) {
|
||||
results = store.getObservationsByIds(rankedIds, { limit: filters.limit || 20 });
|
||||
// Restore semantic ranking order
|
||||
results.sort((a, b) => rankedIds.indexOf(a.id) - rankedIds.indexOf(b.id));
|
||||
}
|
||||
}
|
||||
} catch (chromaError: any) {
|
||||
console.error('[search-server] Chroma ranking failed, using SQLite order:', chromaError.message);
|
||||
// Fall through to SQLite fallback
|
||||
}
|
||||
}
|
||||
|
||||
// Fall back to SQLite-only if Chroma unavailable or failed
|
||||
if (results.length === 0) {
|
||||
console.error('[search-server] Using SQLite-only concept search');
|
||||
results = search.findByConcept(concept, filters);
|
||||
}
|
||||
|
||||
if (results.length === 0) {
|
||||
return {
|
||||
@@ -457,9 +633,59 @@ const tools = [
|
||||
handler: async (args: any) => {
|
||||
try {
|
||||
const { filePath, format = 'index', ...filters } = args;
|
||||
const results = search.findByFile(filePath, filters);
|
||||
let observations: ObservationSearchResult[] = [];
|
||||
let sessions: SessionSummarySearchResult[] = [];
|
||||
|
||||
const totalResults = results.observations.length + results.sessions.length;
|
||||
// Metadata-first, semantic-enhanced search for observations
|
||||
if (chromaClient) {
|
||||
try {
|
||||
console.error('[search-server] Using metadata-first + semantic ranking for file search');
|
||||
|
||||
// Step 1: SQLite metadata filter (get all results with this file)
|
||||
const metadataResults = search.findByFile(filePath, filters);
|
||||
console.error(`[search-server] Found ${metadataResults.observations.length} observations, ${metadataResults.sessions.length} sessions for file "${filePath}"`);
|
||||
|
||||
// Sessions: Keep as-is (already summarized, no semantic ranking needed)
|
||||
sessions = metadataResults.sessions;
|
||||
|
||||
// Observations: Apply semantic ranking
|
||||
if (metadataResults.observations.length > 0) {
|
||||
// Step 2: Chroma semantic ranking (rank by relevance to file path)
|
||||
const ids = metadataResults.observations.map(obs => obs.id);
|
||||
const chromaResults = await queryChroma(filePath, Math.min(ids.length, 100));
|
||||
|
||||
// Intersect: Keep only IDs that passed metadata filter, in semantic rank order
|
||||
const rankedIds: number[] = [];
|
||||
for (const chromaId of chromaResults.ids) {
|
||||
if (ids.includes(chromaId) && !rankedIds.includes(chromaId)) {
|
||||
rankedIds.push(chromaId);
|
||||
}
|
||||
}
|
||||
|
||||
console.error(`[search-server] Chroma ranked ${rankedIds.length} observations by semantic relevance`);
|
||||
|
||||
// Step 3: Hydrate in semantic rank order
|
||||
if (rankedIds.length > 0) {
|
||||
observations = store.getObservationsByIds(rankedIds, { limit: filters.limit || 20 });
|
||||
// Restore semantic ranking order
|
||||
observations.sort((a, b) => rankedIds.indexOf(a.id) - rankedIds.indexOf(b.id));
|
||||
}
|
||||
}
|
||||
} catch (chromaError: any) {
|
||||
console.error('[search-server] Chroma ranking failed, using SQLite order:', chromaError.message);
|
||||
// Fall through to SQLite fallback
|
||||
}
|
||||
}
|
||||
|
||||
// Fall back to SQLite-only if Chroma unavailable or failed
|
||||
if (observations.length === 0 && sessions.length === 0) {
|
||||
console.error('[search-server] Using SQLite-only file search');
|
||||
const results = search.findByFile(filePath, filters);
|
||||
observations = results.observations;
|
||||
sessions = results.sessions;
|
||||
}
|
||||
|
||||
const totalResults = observations.length + sessions.length;
|
||||
|
||||
if (totalResults === 0) {
|
||||
return {
|
||||
@@ -476,13 +702,13 @@ const tools = [
|
||||
const formattedResults: string[] = [];
|
||||
|
||||
// Add observations
|
||||
results.observations.forEach((obs, i) => {
|
||||
observations.forEach((obs, i) => {
|
||||
formattedResults.push(formatObservationIndex(obs, i));
|
||||
});
|
||||
|
||||
// Add sessions
|
||||
results.sessions.forEach((session, i) => {
|
||||
formattedResults.push(formatSessionIndex(session, i + results.observations.length));
|
||||
sessions.forEach((session, i) => {
|
||||
formattedResults.push(formatSessionIndex(session, i + observations.length));
|
||||
});
|
||||
|
||||
combinedText = header + formattedResults.join('\n\n') + formatSearchTips();
|
||||
@@ -490,13 +716,13 @@ const tools = [
|
||||
const formattedResults: string[] = [];
|
||||
|
||||
// Add observations
|
||||
results.observations.forEach((obs, i) => {
|
||||
observations.forEach((obs, i) => {
|
||||
formattedResults.push(formatObservationResult(obs, i));
|
||||
});
|
||||
|
||||
// Add sessions
|
||||
results.sessions.forEach((session, i) => {
|
||||
formattedResults.push(formatSessionResult(session, i + results.observations.length));
|
||||
sessions.forEach((session, i) => {
|
||||
formattedResults.push(formatSessionResult(session, i + observations.length));
|
||||
});
|
||||
|
||||
combinedText = formattedResults.join('\n\n---\n\n');
|
||||
@@ -540,10 +766,53 @@ const tools = [
|
||||
handler: async (args: any) => {
|
||||
try {
|
||||
const { type, format = 'index', ...filters } = args;
|
||||
const results = search.findByType(type, filters);
|
||||
const typeStr = Array.isArray(type) ? type.join(', ') : type;
|
||||
let results: ObservationSearchResult[] = [];
|
||||
|
||||
// Metadata-first, semantic-enhanced search
|
||||
if (chromaClient) {
|
||||
try {
|
||||
console.error('[search-server] Using metadata-first + semantic ranking for type search');
|
||||
|
||||
// Step 1: SQLite metadata filter (get all IDs with this type)
|
||||
const metadataResults = search.findByType(type, filters);
|
||||
console.error(`[search-server] Found ${metadataResults.length} observations with type "${typeStr}"`);
|
||||
|
||||
if (metadataResults.length > 0) {
|
||||
// Step 2: Chroma semantic ranking (rank by relevance to type)
|
||||
const ids = metadataResults.map(obs => obs.id);
|
||||
const chromaResults = await queryChroma(typeStr, Math.min(ids.length, 100));
|
||||
|
||||
// Intersect: Keep only IDs that passed metadata filter, in semantic rank order
|
||||
const rankedIds: number[] = [];
|
||||
for (const chromaId of chromaResults.ids) {
|
||||
if (ids.includes(chromaId) && !rankedIds.includes(chromaId)) {
|
||||
rankedIds.push(chromaId);
|
||||
}
|
||||
}
|
||||
|
||||
console.error(`[search-server] Chroma ranked ${rankedIds.length} results by semantic relevance`);
|
||||
|
||||
// Step 3: Hydrate in semantic rank order
|
||||
if (rankedIds.length > 0) {
|
||||
results = store.getObservationsByIds(rankedIds, { limit: filters.limit || 20 });
|
||||
// Restore semantic ranking order
|
||||
results.sort((a, b) => rankedIds.indexOf(a.id) - rankedIds.indexOf(b.id));
|
||||
}
|
||||
}
|
||||
} catch (chromaError: any) {
|
||||
console.error('[search-server] Chroma ranking failed, using SQLite order:', chromaError.message);
|
||||
// Fall through to SQLite fallback
|
||||
}
|
||||
}
|
||||
|
||||
// Fall back to SQLite-only if Chroma unavailable or failed
|
||||
if (results.length === 0) {
|
||||
console.error('[search-server] Using SQLite-only type search');
|
||||
results = search.findByType(type, filters);
|
||||
}
|
||||
|
||||
if (results.length === 0) {
|
||||
const typeStr = Array.isArray(type) ? type.join(', ') : type;
|
||||
return {
|
||||
content: [{
|
||||
type: 'text' as const,
|
||||
@@ -553,7 +822,6 @@ const tools = [
|
||||
}
|
||||
|
||||
// Format based on requested format
|
||||
const typeStr = Array.isArray(type) ? type.join(', ') : type;
|
||||
let combinedText: string;
|
||||
if (format === 'index') {
|
||||
const header = `Found ${results.length} observation(s) with type "${typeStr}":\n\n`;
|
||||
@@ -827,6 +1095,29 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
|
||||
|
||||
// Start the server
|
||||
async function main() {
|
||||
// Initialize Chroma client
|
||||
try {
|
||||
console.error('[search-server] Initializing Chroma client...');
|
||||
const chromaTransport = new StdioClientTransport({
|
||||
command: 'uvx',
|
||||
args: ['chroma-mcp', '--client-type', 'persistent', '--data-dir', VECTOR_DB_DIR]
|
||||
});
|
||||
|
||||
chromaClient = new Client({
|
||||
name: 'claude-mem-search-chroma-client',
|
||||
version: '1.0.0'
|
||||
}, {
|
||||
capabilities: {}
|
||||
});
|
||||
|
||||
await chromaClient.connect(chromaTransport);
|
||||
console.error('[search-server] Chroma client connected successfully');
|
||||
} catch (error: any) {
|
||||
console.error('[search-server] Failed to initialize Chroma client:', error.message);
|
||||
console.error('[search-server] Falling back to FTS5-only search');
|
||||
chromaClient = null;
|
||||
}
|
||||
|
||||
const transport = new StdioServerTransport();
|
||||
await server.connect(transport);
|
||||
console.error('[search-server] Claude-mem search server started');
|
||||
|
||||
@@ -616,6 +616,33 @@ export class SessionStore {
|
||||
return stmt.all(sdkSessionId) as any[];
|
||||
}
|
||||
|
||||
/**
|
||||
* Get observations by array of IDs with ordering and limit
|
||||
*/
|
||||
getObservationsByIds(
|
||||
ids: number[],
|
||||
options: { orderBy?: 'date_desc' | 'date_asc'; limit?: number } = {}
|
||||
): any[] {
|
||||
if (ids.length === 0) return [];
|
||||
|
||||
const { orderBy = 'date_desc', limit } = options;
|
||||
const orderClause = orderBy === 'date_asc' ? 'ASC' : 'DESC';
|
||||
const limitClause = limit ? `LIMIT ${limit}` : '';
|
||||
|
||||
// Build placeholders for IN clause
|
||||
const placeholders = ids.map(() => '?').join(',');
|
||||
|
||||
const stmt = this.db.prepare(`
|
||||
SELECT *
|
||||
FROM observations
|
||||
WHERE id IN (${placeholders})
|
||||
ORDER BY created_at_epoch ${orderClause}
|
||||
${limitClause}
|
||||
`);
|
||||
|
||||
return stmt.all(...ids) as any[];
|
||||
}
|
||||
|
||||
/**
|
||||
* Get summary for a specific session
|
||||
*/
|
||||
|
||||
@@ -32,6 +32,7 @@ export const TRASH_DIR = join(DATA_DIR, 'trash');
|
||||
export const BACKUPS_DIR = join(DATA_DIR, 'backups');
|
||||
export const USER_SETTINGS_PATH = join(DATA_DIR, 'settings.json');
|
||||
export const DB_PATH = join(DATA_DIR, 'claude-mem.db');
|
||||
export const VECTOR_DB_DIR = join(DATA_DIR, 'vector-db');
|
||||
|
||||
// Claude integration paths
|
||||
export const CLAUDE_SETTINGS_PATH = join(CLAUDE_CONFIG_DIR, 'settings.json');
|
||||
|
||||
Reference in New Issue
Block a user