Which reranker fits your scenario?

Decision guide · ~8 min read ·

Rerankers are not one-size-fits-all. The right choice depends on language coverage, latency budget, data sensitivity, and whether your queries look more like support tickets, legal clauses, code search, or RAG chunks.

In a hurry? The reranker picker asks four questions and gives you a starting point.

Scenario matrix

ScenarioTypical challengeStarting recommendationRun where
RAG pipeline Noisy vector recall, need top-5 precision bge-reranker-v2-m3 or Qwen3-Reranker-4B (self-host), Cohere Rerank 5 or Voyage rerank-3 (API) CPU or GPU / hosted API
Chinese and multilingual Mixed-language queries and documents Qwen3-Reranker, bge-reranker-v2-m3 or gte-multilingual-reranker-base; in mainland China, Alibaba Cloud Model Studio or SiliconFlow Self-host / API in China
Customer support Many similar FAQ lines, fast response jina-reranker-v1-tiny-en or ms-marco MiniLM (English), Cohere Rerank 5 Fast CPU, browser or API
Legal / contracts Subtle wording, distractors share vocabulary An instruction-following reranker: Qwen3-Reranker, Voyage rerank-3, or Contextual AI Rerank v2 (weights for non-commercial use, or its hosted API) Self-host / API
Code & API docs Exact identifiers, syntax matters Hybrid BM25 + vector retrieval, then Qwen3-Reranker or Voyage rerank-3 Hybrid retrieve + rerank
Scanned PDFs, slides, screenshots The text is in an image, or the layout carries meaning Qwen3-VL-Reranker, LightOn-rerank or NVIDIA Nemotron GPU

Each pick links to its page or row, where the vendor’s own figures are shown with how they were measured. None of them replaces a test on your own queries — see evaluating a reranker.

RAG & knowledge bases

Retrieve 50–100 chunks, rerank to 5–10, then prompt. Without a GPU, bge-reranker-v2-m3 is the usual self-hosted default; with one, Qwen3-Reranker-4B, which in Qwen’s own tests beats the 8B on English. As a hosted API, Cohere Rerank 5 and Voyage rerank-3 both bill per token, at the same list price — the cost calculator compares them with the other APIs at your volume. See rerank for RAG and hybrid retrieval when lexical matches matter.

Rule of thumb: if bi-encoder recall is already good, a mid-size reranker is enough. If recall is messy, fix retrieval width first.

Chinese and multilingual

The classic small models — ms-marco MiniLM, Jina v1, mxbai v1 — were trained on English only. Qwen3-Reranker (100+ languages), bge-reranker-v2-m3 and the 306M gte-multilingual-reranker-base (70+ languages) were trained for it. If your servers are in mainland China and you would rather not host a model, Alibaba Cloud Model Studio sells qwen3-rerank by the token in yuan, and SiliconFlow serves bge-reranker-v2-m3 and Qwen3-Reranker behind one endpoint; both are in the models table and the cost calculator’s yuan table. To download open weights in mainland China, hf-mirror.com mirrors Hugging Face. The mainland China guide compares the three routes in detail.

Customer support

Queries are short, passages are FAQ-sized, and latency dominates UX. A tiny cross-encoder (Jina v1 tiny, ms-marco MiniLM) often suffices for English — you are choosing among a handful of similar answers, not reading 100-page PDFs. If MiniLM-L6 is still too slow, its sibling cross-encoder/ms-marco-MiniLM-L4-v2 drops two layers: 73.04 against 74.30 NDCG@10 on TREC DL 19, at 2,500 documents a second against 1,800 on a V100, by sentence-transformers’ own figures. For other languages, start with bge-reranker-v2-m3. Try the Customer support preset in our demo.

Distractors share vocabulary (“termination”, “notice period”, “breach”). Bi-encoders confuse them; cross-encoders separate nuance, and an instruction-following reranker lets you say what counts — “the clause that governs termination for convenience”, “later amendments supersede earlier ones”. Qwen3-Reranker and Voyage rerank-3 take instructions; Contextual AI Rerank v2’s weights are non-commercial. Budget extra latency for a larger model and keep passages to single clauses. Demo preset: Legal clause.

Code & technical docs

Vector search alone misses exact function names; BM25 alone misses paraphrased “how do I…” questions. Use hybrid retrieval, then rerank. Qwen publish code-retrieval scores for Qwen3-Reranker (see the Qwen page), and Voyage report rerank-3’s gain over rerank-2.5 on their code datasets (see the Voyage page). Demo presets: Technical docs and Code search.

Scanned PDFs, slides and screenshots

If your documents are page images — scans, slide decks, dashboards — a text reranker only sees whatever OCR extracted. Qwen3-VL-Reranker (2B / 8B) and LightOn-rerank score page images directly, both under Apache 2.0, and Alibaba Cloud sells qwen3-vl-rerank as an API. NVIDIA’s Nemotron-3.5-Rerank-8B and the smaller llama-nemotron-rerank-vl-1b-v2 do the same under licences that allow commercial use, and take a page’s image and its extracted text together. Their published scores are on ViDoRe, a visual-document benchmark, so compare them with each other rather than with the BEIR column.

Try it in the demo

Each scenario below maps to a built-in preset — load it, hit Rerank, and watch the bi-encoder proxy column mis-order before the cross-encoder fixes it.

Still deciding?

Compare every model and hosted API in one table, then jump into the demo with one click.

Compare models →

Keep reading