Two-Stage Retrieval Pipeline

Bi-encoder recall then cross-encoder precision, before grading and answering

question
route
resolve_context
coref with history
retrieve
Stage 1: bi-encoder
top-N (N = 10)
cheap RECALL, entire KB
rerank_node
Stage 2: cross-encoder
top-N to top-K (K = 3)
focused PRECISION
grade
uses bi-encoder score
answer
terminal router process