A reranker for legal and administrative document search, built on Qwen3-0.6B.
- Parameters
- 609M
- License
- cc-by-nc-4.0
- Max sequence length
- 12,288 tokens
- Updated
- Sep 29, 2026
Search foundation
Reranker
The precision stage of retrieval. A reranker reads the query together with the candidates your search returned and reorders them, so the passages that actually answer the question come first — an upgrade for the search you already run, with no re-indexing.
Released cheon-reranker-0.6b-v1
Playground
Try it on your own text.
Scores come from cheon-reranker-0.6b-v1 running on Hugging Face. Higher means more relevant; scores rank the candidates against each other and are not probabilities. The form starts with the example from the model card.
What is a reranker?
First-stage retrieval — vector search or BM25 — is built for speed. It compares a query with documents that were indexed without ever seeing that query, so the right passage is often in the results but not at the top.
A reranker is the second stage. It takes the query and the retrieved candidates together, judges how well each candidate answers the query, and returns them in a new order. An LLM reads only the top few results, so that order decides what the answer is built on.
How it works
Reranker in 3 steps.
-
01
Retrieve
Your search system — embeddings, BM25 or both — fetches a broad set of candidates quickly.
-
02
Rerank
The reranker reads the query with the candidates and scores how well each one answers it.
-
03
Answer
The best-scored passages go first, to your users or into the prompt of your LLM.
Models
Reranker models
Compare
Reranker, vector search and BM25
| Reranker | Vector search | BM25 | |
|---|---|---|---|
| Stage | Second: reorders candidates | First: retrieves candidates | First: retrieves candidates |
| Reads the query with the text | Yes | No — compares precomputed vectors | No — matches terms |
| Index | None needed | Vector index | Inverted index |
| Cost per query | Highest — a model reads every candidate | Low | Lowest |
| Strongest at | Fine-grained relevance and intent | Meaning and paraphrase | Exact terms, names and codes |
FAQ
Reranker, explained.
Does a reranker replace my search?
No. Retrieval still decides which documents are considered; the reranker reorders what retrieval returns. Keep your index as it is and add reranking on top of its results.
What is listwise reranking?
A pointwise reranker scores each query–candidate pair on its own. A listwise reranker reads the query and the whole candidate list in one pass, so each candidate is judged against the others as well as against the query.
What are the input limits?
cheon-reranker-0.6b-v1: max sequence length 12,288 tokens; max candidates per context 50.
How do I get access?
Released models are on Hugging Face with runnable code in their model cards. For a hosted API or commercial licensing, send a request.
Start a conversation
Put Reranker to work
on your search.
Tell us what you retrieve and how you measure quality.
Get in touch →