Hybrid Search

Keyword search misses paraphrases. Vector search misses the exact error code. GNO runs both, fuses them with reciprocal rank fusion, and reranks with a cross-encoder so relevant passages surface first. Latency depends on the corpus, hardware, enabled stages, and whether models are already loaded. When you want to know why a result ranked where it did, --explain shows the stage timings and per-result score components.

Install GNO Read the docs

Use cases: Teams querying one local index instead of N folders and apps; Researchers mixing exact references with open-ended questions; AI workflows that need accurate retrieval from your own documents

What it gives you

Supported filters apply before candidate limits; empty or fewer-than-requested results remain valid
Opt-in typed custom metadata filters for project, status, numeric confidence, and array membership
Keyword search that handles exact terms, quoted phrases, and technical names
Vector search over paraphrases and conceptual queries
Reciprocal rank fusion of BM25 and vector hits, then cross-encoder rerank
Default seed-scoped graph expansion brings in one-hop linked evidence after initial fusion
Reranking that pushes the most relevant passage to the top
Query expansion that finds related terms you didn’t think to search for
Skips query expansion when the initial match is already strong
Tag, date, category, author, and language filters apply when graph expansion is enabled
Clickable terminal links for command-line search workflows
Configurable speed: fast, balanced, or thorough modes
Switching GPU and CPU, Bun version, or thread count keeps the existing vectors when a sampled re-embed reproduces them
A runtime that cannot reproduce them falls back to keyword retrieval with a warning; a separate vector partition is built only after you confirm it