Querying Architecture#

Search blends lexical and semantic signals with quality & recency, then optionally performs RAG to produce a “fit reason”.

flowchart LR
Q[User query + filters] --> L[Lexical BM25]
Q --> S[Semantic ANN]
D[Catalog DB] --> L
X[Vector Index] --> S
L --> H[Hybrid Ranker]
S --> H
H --> R[RAG optional]
R --> O[Search response]
H --> O

Scoring#

score_final = w_sem*sem + w_lex*lex + w_q*quality + w_r*recency
Weights configured at runtime (e.g., sem=0.6, lex=0.4 by default).

Caching#

• ETag/Last-Modified per query/filter tuple.
• Clients should respect 304 Not Modified.

Failure Modes#

• If vector backend is unavailable, we fall back to lexical only.
• If lexical backend is unavailable, semantic only.
• Results include which modes contributed to the final score.