Concept

Hybrid retrieval

Hybrid retrieval runs vector (semantic), graph, and lexical (keyword) search in parallel and fuses the results with Reciprocal Rank Fusion. No single backend dominates — the strongest signal across all three wins, so you get both fuzzy semantic matches and exact keyword hits.

Parallel backends

A single search request fans out to the vector store, the knowledge graph, and lexical search at once, then merges the candidates.

RRF fusion

Reciprocal Rank Fusion combines the per-backend rankings into one ordered list, rewarding results that multiple backends agree on.

Tracing a query

Pass trace_mode: "debug" to a search to see which backend surfaced each hit, the per-hit scores, and the overall latency — useful when tuning retrieval.