Summary

Each layer catches a different kind of question, and the fusion is what makes one query enough.

Vector search finds "improve error handling", BM25 finds an exact error string, the graph finds connected concepts and metadata filters handle tags and status — reciprocal rank fusion merges them so you do not choose in advance. The MCP process stays alive between queries, so the models load once rather than on every search.

What it is

A retrieval server that fuses vector similarity, BM25 keyword matching, knowledge-graph traversal and metadata filtering, running entirely on ONNX Runtime with local models.

What you get

Five tools: search, index, add_document, health, list_sources
Four retrieval layers merged with reciprocal rank fusion — FAISS vectors, SQLite FTS5 BM25, knowledge graph, metadata filters
Cross-encoder reranking through ONNX Runtime, with no PyTorch in the install
Incremental indexing — only nodes and edges affected by a changed file are rebuilt, with cascading deletes
Header-aware chunking that preserves document structure and parent context
Query expansion for acronyms, casing and spacing variants
CPU-only, no GPU, no API keys, no external services

Requirements

Pip install "velocirag[mcp]", then velocirag index ./my-docs and velocirag mcp. Point it at your data directory with VELOCIRAG_DB or the --db argument. Local models total around 80MB; the reranker downloads its model on first use.

Setup effort

One command — pip install "velocirag[mcp]"

VelociRAG