NodeRAG Retrieves Whole Notes, Not Chunks, Using Local Embeddings and Hybrid Search

A developer has detailed the retrieval half of a personal knowledge system called NodeRAG, which stores notes as plain markdown files and indexes them using a locally run embedding model called nomic-embed-text via Ollama. Unlike standard retrieval-augmented generation systems that split documents into smaller chunks before embedding, NodeRAG treats each self-contained note as a single unit for both indexing and retrieval. The system uses a SQLite database with the vec0 extension for vector search ranked by cosine similarity, paired with a separate full-text index to handle exact keyword queries that dense retrieval often misses. Fusing these two retrieval arms — dense vector search and keyword-based full-text search — yielded measurable recall improvements on a public benchmark dataset. The entire system occupies just 94MB on disk, requires no external API calls, and is designed to be fully inspectable by the user without relying on opaque infrastructure.