codebasesearch

Converts code snippets and natural language queries into dense vector embeddings using Jina's code-aware embedding model, then performs approximate nearest neighbor search against a vector database to find semantically similar code blocks regardless of exact syntax matching. Uses cosine similarity scoring to rank results by semantic relevance rather than keyword overlap, enabling searches like 'authentication middleware' to surface relevant patterns across the codebase.

Scans a codebase directory, extracts code files (respecting .gitignore patterns), chunks them into semantically meaningful units, generates embeddings for each chunk via Jina, and stores vectors in LanceDB with metadata (file path, line numbers, language). Supports incremental re-indexing to update only changed files rather than full re-embedding, reducing computational overhead on large codebases.

Exposes code search capabilities as an MCP (Model Context Protocol) server, allowing Claude, other LLMs, and MCP-compatible clients to invoke semantic code search as a tool within their reasoning loops. Implements MCP resource and tool schemas that map natural language queries to vector search operations, enabling LLM agents to autonomously discover and reference code during code generation or debugging tasks.

Automatically detects programming language from file extension or content, applies language-specific parsing to extract logical code units (functions, classes, methods), and generates embeddings for each unit independently. Preserves language context in embeddings by including language-specific keywords and syntax patterns, enabling Jina's model to understand semantic meaning across Python, JavaScript, TypeScript, Java, Go, Rust, and other languages in a unified vector space.

Computes cosine similarity scores between query embeddings and indexed code embeddings, ranks results by similarity score, and filters results based on configurable similarity thresholds. Allows users to tune precision-recall tradeoffs by adjusting minimum similarity scores, enabling strict matching for high-confidence results or relaxed matching for exploratory search.