VelociRAG
A local retrieval-augmented-generation engine that fuses vector similarity, BM25 keyword search, knowledge-graph traversal and metadata filtering, running entirely on ONNX Runtime with no PyTorch or GPU required.
VelociRAG combines four retrieval methods — vector similarity, BM25 keyword matching, knowledge-graph traversal and metadata filtering — through reciprocal rank fusion with cross-encoder reranking. It runs on ONNX Runtime rather than PyTorch, needs no GPU or API keys, and returns warm queries in about 180 milliseconds.
It ships an MCP server for use from Claude Code, Claude Desktop and Cursor, a Python API for embedding, storing and searching documents directly, and a command-line interface. A Unix-socket daemon keeps the ONNX model and FAISS index warm between CLI queries.
It is written in Python and installs with pip, with local embedding models totalling about 80 MB.