deepak-s-2000/code-rag-mcp

code-rag-mcp

code-rag-mcp is a Retrieval Augmented Generation (RAG) MCP server designed for codebases. It uses FastAPI to provide a scalable, context-aware backend for LLM-based coding assistants (such as Copilot), enabling them to search for existing methods and reuse code efficiently when creating new features.

Features

• Codebase Parsing: Automatically parses supported files in a codebase to extract meaningful code chunks (functions/classes) using tree-sitter.
• Chunk Storage: Stores parsed code chunks with metadata in MongoDB for efficient retrieval.
• Vector Embedding & FAISS Indexing: Converts code chunks into vector embeddings and stores them in a FAISS index for fast similarity search.
• API Endpoints: FastAPI routes to parse codebases, retrieve code chunks, manage vectors, and enable semantic code search.
• Contextual Code Search: LLMs can query for similar code chunks to increase reusability and maintainability, improving feature development velocity.

Architecture Overview

1. Codebase Parsing

• File Types Supported: .py, .cpp, .java, .js, .csv, .json
• Parser: CodeBaseParser traverses the codebase directory, using FileParser with tree-sitter grammars for supported languages.
• Chunk Extraction: AST nodes representing functions or classes are extracted as "chunks," retaining code, type, name, file path, and position.

2. Chunk Storage

• MongoDB: Chunks are inserted into a MongoDB collection via TreeSitterChunksDAO, supporting efficient retrieval and CRUD operations.

3. Vector Embedding & FAISS Store

• Embedding: Uses a configurable model (default: all-MiniLM-L6-v2) to encode code chunks into embeddings.
• FAISS Index: VectorStore manages a FAISS HNSW index for fast inner product searches on code embeddings. The index is persisted to disk (app/rag/faiss/data/code_index.faiss).
• Batch Processing: Chunks are embedded and added to FAISS in batches for scalability.

4. API Endpoints

• /parse: Parses the codebase and stores chunks.
• /get-chunks-by-filename: Retrieve all chunks for a specified file.
• /delete-chunks-by-filename and /delete-all-chunks: Manage chunk storage.
• /create-vectors: Create vector embeddings for stored chunks.
• /search-vectors: Search for similar code chunks by semantic similarity.
• /delete-all-vectors: Reset the FAISS index.

5. FastAPI Application

• Middleware & CORS: Supports CORS and timing middleware for performance logging.
• Background Jobs: Uses APScheduler for periodic tasks (e.g., refresh vectors).
• Modularity: All routes are organized with APIRouter for clean integration.

How It Works

1. Parse the Codebase: POST to /parse to scan the codebase and extract code chunks.
2. Create Embeddings: POST to /create-vectors to embed all chunks and index them in FAISS.
3. Semantic Search: POST to /search-vectors to retrieve existing code similar to a query or feature request.
4. Reuse Code: The system surfaces reusable methods/classes, boosting productivity for Copilot and other LLM agents.

Folder Structure

• app/core/CodeBaseParser.py: Codebase traversal and chunk extraction.
• app/helpers/file_parser.py: Language-specific parsing and AST chunking.
• app/db/tree_sitter_chunks_DAO.py: MongoDB chunk management.
• app/rag/vector_embedding.py: Embedding and FAISS index management.
• app/rag/faiss/vector_store.py: FAISS vector store implementation.
• app/api/routes/parse_codebase.py, app/api/routes/rag_api.py: FastAPI endpoints.

Configuration

• config.py: Change application settings, MongoDB URI, accepted file types, embedding model, and FAISS index path.

Usage

Run the FastAPI server and interact via the documented API endpoints to parse codebases, manage vectors, and enable context-aware code search for LLM agents.

---

License: Apache License 2.0 – see for details.