EKOSdocs
Docs / Architecture Learn

Retrieval

BM25, optional vectors, reciprocal-rank fusion, query understanding and the REASON planner.

A question is understood, planned, retrieved with BM25, vectors and graph traversal, fused, and turned into a typed evidence set and cited answer
A question is understood, planned, retrieved with BM25, vectors and graph traversal, fused, and turned into a typed evidence set and cited answer
Piece Role
Query understanding classifies the question (QueryType) and resolves entities
Rules planner compiles a Query Plan (inspectable via ekos_retrieve / --explain)
Lexical BM25 over names and content excerpts; name matches rank first
Vector (opt-in) brute-force cosine over embeddings
Fusion reciprocal-rank fusion (RRF) of lexical and vector results
Rerank (opt-in) LLM rerank of the top candidates
Evidence set typed atomic claims with location and producing analyzer

Retrieval quality has a CI-gated evaluation harness, so ranking regressions fail the build.