Searching the knowledge base
The search panel in the knowledge base runs every query against approved and locked content only. Results are scoped to the current project before they exist: the permission boundary lives in the database query, not in a post-filter step.
Ask Atlas (agentic mode)
✦ Agentic is on by default. When it is active, Atlas receives your question and investigates using semantic search, code browsing, and artifact lookup. It streams a synthesized answer with citations back into the panel. A model selector (Auto by default) sits in the input footer.
Switch ✦ Agentic off to run one of the four direct modes.
Direct search modes
Hybrid is the active mode when you switch out of agentic. The other modes are available from the mode tabs at the top of the panel.
How hybrid retrieval works
Hybrid mode runs two searches in parallel and fuses the results.
The vector search embeds your query and finds the closest chunks by cosine similarity using the pgvector index. The full-text search runs a Postgres ts_vector query against the chunk content. Both sets of results are merged with reciprocal rank fusion (RRF), weighted to favor the vector ranking. The weights and the RRF constant k were tuned against BEIR fiqa, trec-covid, and MultiHop-RAG ablations.
If the vector index is unavailable, the search falls back to an ILIKE query so results still arrive.
Optional reranker
An administrator can configure a reranker to re-score the fused results before they reach you. Four backends are supported:
If any backend returns an error or times out, the reranker fails open: the original fusion order is returned with a warning log, and the search result is unaffected.