Interactive · Week 6 · Section 3

How a reranker fixes mixed retrieval.

Same 60-chunk corpus from Week 5. Stage 1 is embedding retrieval — it pulls a top-20 candidate set by geometric distance, which mixes chunks from adjacent topical clusters. Stage 2 is the reranker — a cross-encoder that reads each chunk together with the query and re-scores by actual content relevance. Watch how the right answers climb from rank 14 to rank 1, and the visually-close-but-irrelevant chunks fall out of the final top-5.

Query

First-stage K 20

Final K 5

Embedding space · candidate retrieval shown geometrically

faded dots aren't candidates · accent rings mark final top-K

stage 1 Embedding retrieval

cheap · top-K1 by vector distance · breadth, not precision

stage 2 Cross-encoder reranker

expensive · re-scores top-K1, returns top-K2 by relevance

What's happening here

Stage 1 score is geometric: max(0, 1 − d/600) on Euclidean distance in the 2D projection — the cheap, vector-comparable baseline. Stage 2 simulates a cross-encoder by combining lexical overlap with a topic-relevance bonus that a real cross-encoder would learn from training data: min(1, base_overlap + cluster_bonus) where cluster_bonus ≈ 0.55 for chunks in the query's primary topic. Real cross-encoders work by encoding query and chunk together through one transformer pass — much more accurate than embedding comparison, much slower per pair, and worth the 200–500 ms latency hit on most production systems. Source-of-truth lives as <RerankerVisualizer/> in the curriculum codebase.