Understand rerankers. Then watch one run in your browser.
A reranker re-scores your retrieved candidates so the most relevant passages rise to the top. Learn how it works, compare the popular models, and try a cross-encoder live — with zero API cost and nothing leaving your machine.
Runs on transformers.js · 100% client-side · no key required
Start here
Three short guides take you from “what is a reranker?” to a working reranking stage in your RAG pipeline.
What is a reranker?
The two-stage retrieval pattern, why order matters, and where reranking fits.
A reranker scores (query, document) pairs jointly — far more accurate than cosine similarity alone.
Cross-encoder vs bi-encoder
Why bi-encoders are fast and cross-encoders are accurate — and how to use both.
Bi-encoders retrieve at scale; cross-encoders rank with precision. Use them in two stages.
How to add reranking to RAG
Retrieve wide, rerank, keep the best. With code, top-k tips and latency trade-offs.
Retrieve 50+ candidates, rerank, keep 5 — better answers and fewer tokens to the LLM.
Jump to the job
Already know what you need? Skip the tour and open the page that answers it.
Choose by scenario
RAG, support, legal, code or Chinese search — which reranker actually fits.
Reranking in mainland China
Alibaba Cloud, SiliconFlow, or your own GPU — prices in yuan, and how to get the weights.
What will it cost?
Cohere per search, Voyage per token, plus Alibaba Cloud and SiliconFlow in yuan.
Reranking didn't help
Seven usual causes — recall, pool size, truncation — and how to tell which one you have.
Rerank in your browser, right now
Paste a query and a few candidate passages. A real cross-encoder downloads once, caches, and scores every pair locally — you watch the ranking reshuffle in milliseconds. No server, no API key, no data leaving the page.
- Real model weights via transformers.js + ONNX Runtime Web
- Zero API cost and zero abuse risk — it’s all on your device
- See exactly how scores reorder your retrieval results
Compare the rerank models
2026 open SOTA candidates (Qwen3-Reranker), Jina v3.5 listwise, classic bge/mxbai, hosted Cohere/Voyage — plus late-interaction and instruction APIs.
Qwen3-Reranker
0.6B / 4B / 8B Apache 2.0 family — the multilingual GPU self-host pick. Start at 4B, not 8B.
bge-reranker
Open-weight rerankers from BAAI. Still the best CPU-friendly multilingual default.
Cohere Rerank 4
Mature hosted API in Pro and Fast variants — 32k context, 100+ languages, billed per search.
Jina Reranker v3.5
0.6B listwise model at 63.20* BEIR in Jina's own test, level with Qwen3-Reranker-4B. Weights are non-commercial; v1-tiny powers our demo.
Voyage rerank-3
Hosted rerankers you steer with a natural-language instruction. 32k context, per-token pricing.
mxbai-rerank
Apache 2.0 open weights. v1's xsmall runs in the browser; v2 adds 100+ languages, including Chinese.
Also: ColBERT / late-interaction · instruction-following rerank
Reranking in one diagram
- Input Query
- fast · recall Retriever bi-encoder / BM25 returns 50–100
- slow · precise Reranker cross-encoder scores keeps the top 5
- Answer LLM generate