Code Vector Search (MCP Server)

A Model Context Protocol (MCP) server that builds a local semantic code index using ChromaDB and sentence-transformer embeddings accelerated with Intel OpenVINO. It watches your project, chunks source files, respects .gitignore, and exposes semantic search tools to MCP clients.

Features

• ChromaDB-backed vector store with persistent storage and automatic bloat mitigation.
• Sentence-transformer embeddings with OpenVINO acceleration for Intel CPUs/GPUs (falls back to PyTorch).
• Gitignore-aware indexing with file watching for live updates.
• Chunked documents for precise results with file/line metadata.
• Tools exposed to MCP clients: semantic_search, get_index_stats, reindex_project.

Quickstart (Docker)

docker build -t code-vector-search .
PROJECT_ROOT=/path/to/your/code ./run-docker.sh

• Project is mounted read-only at /project; the vector index persists to /path/to/your/code/.vector-index.
• Set DEBUG=1 to surface server stderr logs from the container.

Running without Docker

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
PROJECT_ROOT=/path/to/your/code PERSIST_DIR=.vector-index python server.py

Configuration

• PROJECT_ROOT (default /project): directory to index.
• PERSIST_DIR (default /vector-index): where ChromaDB stores the index.
• FORCE_REINDEX (true/false): force a clean rebuild on startup.
• CHUNK_SIZE (default 800): target characters per chunk; larger values reduce chunks and speed indexing.
• EMBED_BATCH_SIZE (default 64): sentence-transformer encode batch size.
• EMBED_MODEL_NAME (default all-MiniLM-L6-v2): embedding model name for SentenceTransformers.
• INDEX_BATCH_CHUNKS (default 512): number of chunks per Chroma add batch during full indexing.
• DEBUG (with run-docker.sh): surface stderr logs instead of silencing them.

Performance tips

• These defaults are tuned for a 4-core, 32GB machine; adjust down if you see memory pressure.
• Increase EMBED_BATCH_SIZE (e.g., 96) if you have enough headroom; lower it if you see OOMs.
• For very large repos, prefer a fresh index after big changes to avoid HNSW bloat.
• Model choices: all-MiniLM-L12-v2 (higher quality, slower), multi-qa-MiniLM-L6-cos-v1 (QA-style queries), bge-small-en-v1.5 (faster, smaller).

Tests

Run embedding and Chroma integration checks:

python test_embedding.py

Contributing

See CONTRIBUTING.md for setup, style, and pull request guidelines. A CODE_OF_CONDUCT.md applies to all community spaces.

Security

Please review SECURITY.md for how to report vulnerabilities.

License

Licensed under the MIT License. See LICENSE for details.