MCP Semantic Search

A Model Context Protocol (MCP) server that indexes codebases using semantic embeddings for natural language search.

Features

• 🔍 Semantic Code Search – Find code using natural language queries instead of exact text matching
• ⚡ Fast Indexing – Efficient chunking and batch embedding with background processing
• 🧠 Smart Chunking – Language-aware code splitting:
  • Python: Function/class boundary detection
  • Others: Line-based with configurable overlap
• 🌐 Multi-language Support – Python, JavaScript, TypeScript, JSX, TSX, Markdown, YAML, JSON, HTML, CSS, Bash, SQL, and more
• 👀 Live Watch – Automatically re-index on file changes with debouncing
• 🔄 Incremental Updates – Reindex only changed files without full rebuild
• 🗑️ Deletion Handling – Automatically removes chunks for deleted files
• 📊 Status Tracking – Real-time indexing progress and queue monitoring

Quick Start

Prerequisites

• Python 3.12 or higher
• Qdrant vector database (running locally or remotely)
• Google Gemini API key

Installation

# Using uvx (recommended - no installation needed)
uvx mcp-semantic-search

# Or install with pip
pip install mcp-semantic-search

Configuration

Set environment variables:

export GEMINI_API_KEY="your_gemini_api_key"
export QDRANT_URL="http://localhost:6333"

Optional environment variables:

# Embedding model (default: text-embedding-004)
export GEMINI_EMBEDDING_MODEL="text-embedding-004"

# Chunk configuration (defaults: 50/10/5)
export CHUNK_MAX_LINES=50        # Max lines per chunk
export CHUNK_OVERLAP_LINES=10    # Overlap between chunks
export CHUNK_MIN_LINES=5         # Min lines for valid chunk

Or create a .env file:

GEMINI_API_KEY=your_gemini_api_key
QDRANT_URL=http://localhost:6333

Running Qdrant

# Using Docker
docker run -p 6333:6333 qdrant/qdrant

# Or using docker-compose
echo '
services:
  qdrant:
    image: qdrant/qdrant
    ports:
      - "6333:6333"
' | docker-compose -f - up

Usage with Claude Code

Method 1: Using MCP Config JSON (Recommended)

Edit your Claude Code MCP configuration file (~/.claude.json or ~/.config/claude/config.json):

{
  "mcpServers": {
    "semantic-search": {
      "type": "stdio",
      "command": "uvx",
      "args": ["mcp-semantic-search"],
      "env": {
        "GEMINI_API_KEY": "your_gemini_api_key_here",
        "QDRANT_URL": "http://localhost:6333"
      }
    }
  }
}

For a local installation (after pip install mcp-semantic-search):

{
  "mcpServers": {
    "semantic-search": {
      "type": "stdio",
      "command": "mcp-semantic-search",
      "env": {
        "GEMINI_API_KEY": "your_gemini_api_key_here",
        "QDRANT_URL": "http://localhost:6333"
      }
    }
  }
}

With optional chunk configuration:

{
  "mcpServers": {
    "semantic-search": {
      "type": "stdio",
      "command": "uvx",
      "args": ["mcp-semantic-search"],
      "env": {
        "GEMINI_API_KEY": "your_gemini_api_key_here",
        "QDRANT_URL": "http://localhost:6333",
        "CHUNK_MAX_LINES": "50",
        "CHUNK_OVERLAP_LINES": "10",
        "CHUNK_MIN_LINES": "5"
      }
    }
  }
}

Method 2: Using CLI

claude mcp add semantic-search \
  -e GEMINI_API_KEY="$GEMINI_API_KEY" \
  -e QDRANT_URL="$QDRANT_URL" \
  -- uvx mcp-semantic-search

Available Tools

Tool	Description	Returns
index_codebase(root_dir, force_reindex, max_files)	Index the codebase	{"status": "success", "files_queued": N}
search_code(query, limit, score_threshold)	Semantic search across all files	{"query": "...", "count": N, "results": [...]}
search_file(query, file_path, limit)	Search within a specific file	{"query": "...", "file": "...", "results": [...]}
get_status()	Check indexing status	{"collection": {...}, "queue": {...}}
start_live_watch(root_dir, debounce_seconds)	Start file watching	{"status": "success", "running": true}
stop_live_watch()	Stop file watching	{"status": "stopped", "running": false}
clear_index()	Reset the entire index	{"status": "success", "message": "..."}

Example Workflow

# Index your codebase (auto-starts on first use)
index_codebase(root_dir="/path/to/project")
# Returns: {"status": "success", "files_queued": 1234}

# Search for code using natural language
search_code("how does authentication work")
# Returns:
# {
#   "query": "...",
#   "count": 5,
#   "results": [
#     {
#       "file": "src/auth/middleware.py",
#       "lines": "10-25",
#       "score": 0.876,
#       "content": "..."
#     },
#     ...
#   ]
# }

# Check indexing status
get_status()
# Returns:
# {
#   "collection": {"total_chunks": 12345, "files_indexed": 1234},
#   "queue": {"running": true, "queued": 0, "pending": 0}
# }

# Enable live watching (auto-index on file changes)
start_live_watch(root_dir="/path/to/project")

Configuration

Chunking Configuration

Control how code is split into searchable chunks:

# Smaller chunks = more precise results, more storage
export CHUNK_MAX_LINES=30

# Larger chunks = more context per result
export CHUNK_MAX_LINES=100

# Adjust overlap for context continuity
export CHUNK_OVERLAP_LINES=15

Variable	Default	Description
CHUNK_MAX_LINES	50	Maximum lines per chunk
CHUNK_OVERLAP_LINES	10	Overlap between chunks
CHUNK_MIN_LINES	5	Minimum lines for valid chunk

Search Configuration

# Adjust search parameters
search_code(
    query="your query",
    limit=20,                 # More results (default: 10)
    score_threshold=0.3       # Lower threshold = more results (default: 0.5)
)

Development

Setup

# Clone the repository
git clone https://github.com/yourusername/mcp-semantic-search.git
cd mcp-semantic-search

# Install in development mode
pip install -e .

Testing

# Test with a small subset
python -c "
from mcp_semantic_search import GeminiEmbedder, QdrantCodeStore, index_repository

embedder = GeminiEmbedder()
store = QdrantCodeStore()

# Test with just 5 files
stats = index_repository(
    root_dir='.',
    embedder=embedder,
    store=store,
    max_files=5
)
print(stats)
"

# Test semantic search
python -c "
from mcp_semantic_search import GeminiEmbedder, QdrantCodeStore

embedder = GeminiEmbedder()
store = QdrantCodeStore()

query_embedding = embedder.embed_query('authentication')
results = store.search(query_embedding, limit=5)

for r in results:
    print(f'{r[\"file_path\"]}:{r[\"start_line\"]} ({r[\"score\"]:.2f})')
    print(r['content'][:200])
    print('---')
"

Technical Details

• Embedding Model: Google text-embedding-004 (768 dimensions)
• Vector Database: Qdrant with cosine similarity
• Chunking Strategy:
  • Python: AST-based function/class boundary detection
  • Others: Line-based with configurable chunk size and overlap
• File Watching: Watchdog with 3-second debouncing
• Deduplication: SHA256 hash-based, unchanged files are skipped
• Background Processing: FIFO queue for incremental reindexing

Supported Languages

Extension	Language
.py	Python
.js	JavaScript
.ts	TypeScript
.jsx	JSX
.tsx	TSX
.md	Markdown
.yaml, .yml	YAML
.json	JSON
.html	HTML
.css	CSS
.sh	Bash
.sql	SQL
.txt	Text

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgments

• Model Context Protocol by Anthropic
• Qdrant - Vector Database
• Google Gemini - Embedding API
• fastmcp - MCP framework