Extraction
Pull hallucination-free answers straight from the source
Overview
Extraction
Extractive question answering, or answer extraction, is the process of extracting answers to questions from documents — for example, “what is the governing law of this contract?” or “who is the defendant in this case?”. Extractive QA can solve a diverse range of information extraction problems, from pulling out key dates to highlighting legal doctrines relied upon in a judgment.
Use cases
Benchmark-leading accuracy
The world’s best legal extractive QA models of their size.
Kanon Answer Extractor and Kanon Answer Extractor Mini outperform OpenAI’s GPT-4.1 by over 90% on LegalQAEval — the only extractive question answering benchmark set for the legal domain.
As extractive question answering models, they excel at pulling out answers to questions from legal documents, whether it is as simple as “what is the governing law of this contract?” or as complex as “what doctrine did the judge rely on in making her decision?”.
Source-grounded answers
Answers come directly from your inputs — immune to generative hallucinations.
Some legal products do not need generated prose. They need the answer, exactly where it appears in the source. Kanon Answer Extractor reads the input document and pulls out the answer to a question directly from that text — it does not invent an answer.
Extractive QA can often prove faster, more precise, and more cost-effective than generative QA by virtue of being specialized for a single, domain-specific task.
Purpose-built over generic
Specialized for extracting legal answers from legal text — not paraphrasing or inferring.
Extraction is useful anywhere your product needs a short, reliable answer from a long legal document — contract review, due diligence, case analysis, regulatory review, evidence workflows, and internal knowledge systems.
Instead of asking a general-purpose model to summarize or infer, send a document and a question, get back the answer span, and build the surrounding product experience around source-grounded outputs.