@13w/local-rag

Implements a distributed semantic memory layer using Qdrant vector database as the backend storage, enabling Claude Code agents to persist and retrieve embeddings across sessions. The system stores embeddings generated from code snippets, documentation, and conversation context in a vector index, allowing agents to maintain long-term semantic understanding without re-embedding identical content. Uses MCP protocol to expose memory operations as standardized tools that Claude can invoke during code generation and reasoning tasks.

Provides semantic search over codebases by generating embeddings that incorporate code structure awareness, not just raw text similarity. The system can index code files, extract meaningful code units (functions, classes, modules), and generate embeddings that capture both semantic meaning and syntactic context. Search queries return ranked code snippets with relevance scores, enabling Claude agents to find relevant code patterns and implementations without keyword matching.

Wraps all RAG and memory operations as MCP (Model Context Protocol) tools that Claude Code agents can invoke directly, using MCP's standardized tool schema and request/response format. The system registers tools for memory operations (store, retrieve, search, delete) and exposes them through the MCP server interface, allowing Claude to autonomously decide when to access memory without requiring custom prompt engineering or wrapper code.

Integrates with Ollama to generate embeddings locally without external API calls, using open-source embedding models (e.g., nomic-embed-text, all-minilm). The system can invoke Ollama's embedding endpoint to convert code snippets and search queries into vector representations, enabling fully local RAG pipelines without dependency on commercial embedding APIs. Supports fallback to external embedding APIs if Ollama is unavailable.

Supports indexing and semantic search across multiple programming languages (JavaScript, TypeScript, Python, Go, Rust, etc.) by using language-agnostic embedding generation and optional language-specific parsing for code structure awareness. The system can index mixed-language codebases, maintain separate vector indices per language if needed, and retrieve relevant code regardless of language boundaries. Enables cross-language code pattern discovery and reuse.

Stores embeddings with rich metadata (file paths, function signatures, timestamps, code language, author, etc.) and enables filtering/retrieval based on metadata predicates, not just semantic similarity. The system can retrieve embeddings matching specific criteria (e.g., 'all Python functions modified in last week', 'all code in src/utils directory') and combine metadata filtering with semantic search for precise context retrieval. Metadata is stored alongside vectors in Qdrant using payload filtering.

Provides memory isolation mechanisms that allow different Claude Code agents or sessions to maintain separate memory spaces, preventing cross-contamination of context. The system can scope memory operations to specific sessions, users, or projects using namespace/partition strategies in Qdrant, enabling multiple agents to operate independently while sharing the same vector database infrastructure. Supports both isolated and shared memory modes depending on use case.

Supports incremental indexing of codebase changes rather than full re-indexing, using file modification timestamps or git diff to detect changed files and update only affected embeddings. The system can track which files have been indexed, detect changes since last indexing, and update only the changed code units in the vector database. Enables efficient maintenance of large codebase indices without full re-embedding on every update.

Supports hybrid search that combines semantic vector similarity with keyword/BM25 matching, enabling retrieval that balances semantic understanding with exact term matching. The system can execute both semantic and keyword searches in parallel, rank results using combined scores, and return results that capture both semantic relevance and keyword precision. Useful for code search where exact function names or identifiers matter alongside semantic similarity.