𝐇𝐨𝐰 𝐭𝐨 𝐑𝐞𝐫𝐚𝐧𝐤 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥 𝐑𝐞𝐬𝐮𝐥𝐭𝐬 𝐰𝐢𝐭𝐡 𝐚 𝐂𝐫𝐨𝐬𝐬-𝐄𝐧𝐜𝐨𝐝𝐞𝐫

Your bi-encoder retrieves fast but ranks poorly. Your cross-encoder ranks well but is too slow to search the whole corpus. So you run both - retrieve 20 candidates, rerank them, measure the NDCG@10 gain yourself on your own machine.

Most retrieval pipelines stop at the retriever. They don't measure what they're leaving on the table. This guide walks you through building a two-stage pipeline, computing NDCG@10 before and after reranking, and watching the delta move from 0.8447 to 0.9228 - a +7.8% jump in ranking quality. On CPU. No API keys.

You'll learn why bi-encoders and cross-encoders answer different questions. A bi-encoder compresses each passage into a vector ahead of time - fast, but the query and passage never meet. A cross-encoder runs a transformer over both together - slower per pair, but it sees the interaction. The reranker fixes what the retriever got wrong: passages that share vocabulary but answer a different question.

The tutorial covers the core pattern: build a small corpus with planted decoys (passages the bi-encoder ranks too highly), retrieve candidates, score them with the cross-encoder, compute NDCG@10 on both rankings, measure the latency tradeoff. Everything here runs on CPU with ~180 MB of model downloads. Intermediate level, about 30 minutes. You'll finish with two numbers you generated yourself - not a vendor benchmark.