The first results are fast. They are not the best.

Two-stage retrieval: cast a wide cheap net, then re-score the survivors with a slow accurate model.

Warm-up (Lesson 2 recall): why merge dense and sparse with RRF instead of adding their scores? Because the score scales do not match, so RRF merges on rank position. Said it without peeking? On to reranking.

The one idea: Stage-3 retrieval has two passes. Pass one is built for speed and gets you the top ~100 roughly-right chunks. Pass two is built for accuracy and re-orders just those 100 so the truly-best float to the top 10 the LLM actually reads.

Why two passes exist: bi-encoder vs cross-encoder

The thing that makes vector search fast is also what limits it. It is a bi-encoder: it turns the query into a vector and each document into a vector separately, then compares them. Documents are embedded once at index time, so query time is a near-instant nearest-neighbor lookup. Fast, scales to millions, but the query and document never actually "see" each other.

A cross-encoder does the opposite. It feeds the query and one document through the model together, so every query word can attend to every document word. Far more accurate at judging "does this actually answer the question." But it cannot precompute anything: it runs a full model pass for every (query, document) pair. Too slow to run over your whole store, perfect to run over 100 candidates.

	Bi-encoder (pass 1)	Cross-encoder (pass 2)
Sees query+doc together?	No, encoded separately	Yes, jointly
Speed	Instant lookup, precomputed	Slow, one pass per pair
Runs over	The whole store (millions)	A short candidate list (~100)
Job	High recall: don't miss the answer	High precision: order it right

Recall = "the answer is somewhere in the 100." Precision = "the answer is in the top few." Pass 1 buys recall cheaply; pass 2 buys precision on a small set. Running cross-encoder precision at bi-encoder scale would cost a fortune, so you don't.

The pattern, end to end

Hybrid (L2)dense + sparse + RRF → top ~100

Rerankcross-encoder scores each pair

Top 10precision-tuned, sent to the LLM

SynthesizeLLM answers from the best 10

The payoff is real and measured: adding a cross-encoder reranker typically lifts a RAG system from ~65-80% relevance to ~85-90% on hard queries, for +5 to +15 NDCG points. It is the single highest-leverage upgrade after you already have hybrid search.

Connect it back: gbrain's default embedder/reranker is ZeroEntropy (per its README). Now you know what that reranker is doing: the pass-two cross-encoder step on this diagram. "Search gives you raw pages, GBrain gives you the answer" leans on exactly this precision pass before synthesis.

Retrieval practice

Why not just run the cross-encoder over the entire store?

What is the job of the fast first pass in two-stage retrieval?

What can a cross-encoder do that a bi-encoder cannot?

Read this next (primary source)

★ Bi-Encoders vs Cross-Encoders (ZeroEntropy, the team behind gbrain's default reranker). The clearest write-up of today's lesson. ~10 min.

Also good: Advanced RAG Retrieval: Cross-Encoders & Reranking (TDS).

Ask your teacher. Try: "does my Graphiti setup rerank, or does it skip straight to graph traversal?", "what reranker should a Rejig client KB use, open-source or API?", or "if reranking is so good, why isn't it always on?" (latency and cost — worth discussing for client builds).

← Lesson 2: Meaning and exact match Next → Lesson 4: Chunking and embeddings (garbage in, garbage out)