Guides
From “what is a reranker?” to hybrid retrieval pipelines and evaluation — pick a starting point or read straight through.
Read these in order
Five guides that build on each other, from what a reranker does to measuring one on your own data.
-
What is a reranker?
Two-stage retrieval, why order matters, when reranking pays off.
Fundamentals~7 min -
Cross-encoder vs bi-encoder
Speed vs accuracy — and why production stacks use both.
Architecture~8 min -
How to add reranking to RAG
Retrieve wide, rerank tight — with Python, LangChain, and latency tips.
Practical~10 min -
Hybrid retrieval + rerank
Combine BM25 and vector search, then let the reranker fix the merge.
Production~9 min -
How to evaluate rerankers
NDCG, MRR, labelled queries, and regression tests for your domain.
Evaluation~8 min
Already have a specific question?
Go straight to the page for the job in front of you.
Choose by scenario
RAG vs support vs legal vs code — which model fits your use case.
Compare the rerankers
Qwen3, bge, mxbai, Jina, GTE and the hosted APIs, set side by side only where the numbers come from one run.
Reranking in mainland China
Alibaba Cloud, SiliconFlow or your own server — prices in yuan, Chinese scores, downloading weights.
Reranking didn't help
Seven reasons you see no lift — recall, pool size, truncation — and how to tell which one you have.
Rerank cost calculator
Per-search or per-token? Which rerank API is cheaper at your volume.
Self-host a reranker
sentence-transformers, GPU sizing, and a minimal FastAPI service.
Rerank on a vector database
Wiring a rerank stage onto pgvector, Qdrant or Elasticsearch, with code.
Late-interaction (ColBERT)
When ColBERT beats bi-encoders and when you still need cross-encoder rerank.
Instruction-following rerank
Task instructions that change what “relevant” means — vs plain cross-encoders.
Live in-browser demo
Watch a cross-encoder reorder passages — bi-encoder proxy included.