Reranking
Put the most relevant legal context first
Overview
Reranking
Reranking is the process of scoring and sorting documents based on their relevance to a query. It is often used with a search engine or embedding retriever: inexpensive first-stage retrieval finds candidate documents, then a reranker resorts them by true relevance before results are shown to the user or passed into a retrieval-augmented generation (RAG) pipeline.
Use cases
Benchmark-leading accuracy
Ranks first on Legal RAG Bench.
As the first reranker optimized specifically for legal RAG, Kanon 2 Reranker excels at scoring the relevance of laws, decisions, contracts, evidence, and other legal documents to legal queries.
When paired with Kanon 2 Embedder as a fast, affordable, and accurate first-stage retriever, Kanon 2 Reranker delivers 18% better legal information retrieval performance on Legal RAG Bench and 6% better performance on the Massive Legal Embedding Benchmark (MLEB).
On Legal RAG Bench, Kanon 2 Reranker is also 7% more accurate than Voyage Rerank 2.5, 9% more accurate than Qwen 3 Reranker 8B, and 24% more accurate than Voyage 4 Large.
NDCG@10 model score bar chart comparing rerankers and embedders on Legal RAG Bench 
Kanon 2
RerankerReranker
Voyage Rerank
2.5Reranker
Qwen Reranker
8BReranker Kanon 2
EmbedderEmbedder Voyage 4
LargeEmbedderPurpose-built over generic
Built for laws, decisions, contracts, and evidence — not generic web text.
Most RAG failures start before generation. The system retrieves something plausible, but not quite the right authority, clause, exhibit, or passage. Reranking gives your retrieval pipeline a second pass before the model starts writing.
Kanon 2 Reranker scores candidate documents against the user’s query and reorders them by true legal relevance. Use it after embedding search, keyword search, hybrid search, or another candidate generator to decide which results deserve to be shown, cited, summarized, or passed into a model.
Documents of any length
Supports documents of any length via semchunk.
A contract clause may need its definitions, a case may need its facts and holding, and evidence may need surrounding context before it can be judged properly.
Uniquely, unlike its closest competitors, Kanon 2 Reranker boasts first-class support for documents of any length powered by semchunk, our semantic chunking library. That lets your application work with long contracts, decisions, filings, bundles, policies, and matter documents while still ranking the pieces that matter most.