Fig. 1 ·
The minimal architecture
two loops · one index
Online · request path
Offline · index path
Query
Retrieve
dense (meaning)
+ sparse (words)
Fuse · Rerank
RRF → cross-encoder
~150 → top 5–10
Generate
grounded answer
Answer
Docs
Chunk
on meaning
Enrich
context · keywords · metadata
Index
vector + lexical
served at query time
control / data flow
retrieval signal
the LLM appears once · at the very end