Introducing Kanon 2 Reranker

tl;dr

We’re releasing Kanon 2 Reranker, the most powerful reranking model for legal RAG to date.

Kanon 2 Reranker ranks first on Legal RAG Bench, outperforming Qwen 3 Reranker 8B by 9% and Voyage Rerank 2.5 by 7%.

Uniquely, Kanon 2 Reranker supports an infinite context window thanks to our semchunk semantic chunking library.

Kanon 2 Reranker is available for use today via the Isaacus API, priced the same as Kanon 2 Embedder at US$0.35 per million tokens.

16 March 2026: Kanon 2 Reranker can now be self-hosted via the AWS Marketplace.

Isaacus Voyage Qwen

Kanon 2
Reranker

Reranker

Voyage Rerank
2.5

Reranker

Qwen Reranker
8B

Reranker

Kanon 2
Embedder

Embedder

Voyage 4
Large

Embedder

Kanon 2 Reranker is the most powerful reranker for legal RAG

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 queries in legal RAG applications.

When paired with our state-of-the-art legal embedding model, 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 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.

For a practical example of how better retrieval translates into real-world improvements in legal RAG quality, consider the below example, where Kanon 2 Reranker enables GPT-5.2 to correctly answer a question that it fails to answer when paired with Voyage Rerank 2.5.

Question

Chris is on trial for intentionally or recklessly causing a bushfire. His counsel says he was carrying out a “cool burn.” Talia, a First Nations woman, is called to explain First Nations practices. Is Talia’s evidence admissible?

Kanon 2 Reranker

Isaacus

Yes. Expert opinion evidence is admissible if “wholly or substantially based on specialised knowledge” from training, study or experience (Evidence Act 2008 s 79), and the bench book also recognises an exception for “evidence about the existence or content of Aboriginal and Torres Strait Islander traditional laws and customs (s 78A)” (DOC: 4.13.1-c2-s1).

Retrieved evidence

#1

4.13.1 Admissibility of Opinion Evidence

Answer

4.13.1-c2-s1

#2

4.13.1 General Principles of Opinion Evidence

4.13.1-c1-s2

#3

4.17 Admissibility of Evidence Capable of Proving a Tendency

4.17-c3-s1

#4

4.19 Determining the Relevance of Evidence

4.19-c3-s1

#5

4.13.1 General Principles of Opinion Evidence

4.13.1-c1-s3

Rerank 2.5

Voyage

No. The provided charge book material for intentionally or recklessly causing a bushfire only sets out the three elements and the relevant states of mind (“aware that … would probably cause a fire/spread”) and contains no rule making cultural practice evidence admissible (DOC: 7.5.21.1-c1-s1; DOC: 7.5.21.1-c2-s1; DOC: 7.5.21.2-c1-s1).

Retrieved evidence

#1

7.5.21.1 Charge: Intentionally or Recklessly Causing a Bushfire

7.5.21.1-c1-s1

#2

7.5.21.2 Checklist: Intentionally or Recklessly Causing a Bushfire

7.5.21.2-c1-s1

#3

7.5.21 Elements

7.5.21-c2-s1

#4

7.5.21.1 Intentionally or recklessly causing a fire

7.5.21.1-c2-s1

#5

7.5.20.1 Intention to Damage or Destroy Property by Fire

7.5.20.1-c3-s1

A side-by-side comparison of how Kanon 2 Reranker performs on Legal RAG Bench relative to Voyage Rerank 2.5 in an end-to-end legal RAG pipeline, with GPT-5.2 used to generate answers.

Uniquely, unlike its closest competitors, Kanon 2 Reranker boasts first-class support for documents of any length powered by our semchunk semantic chunking library.

We have made Kanon 2 Reranker publicly available through the Isaacus API at the same price as Kanon 2 Embedder: US$0.35 per million tokens.

To get started building your own awesome legal AI applications with Kanon 2 Reranker, check out our usage guide.

If you’d like to be updated as we release new state-of-the-art legal models, join our platform and then follow us on LinkedIn, X/Twitter, or Reddit.