Skip to main content
Isaacus offers some of the world’s most accurate, efficient and scalable legal AI models. This page provides an overview of those models, including their capabilities and constraints. If you’re looking to understand the pricing of these models, please refer to our pricing page. If you want to use these models in practice, you can check out our quickstart guide.

​
Embedding

Embedding or vectorization is the process of converting content into sets of numbers that, when compared mathematically with each other, quantify how similar they are in meaning. Embeddings are used to power semantic search engines, text classification, and cluster analysis as well as the retrieval component of retrieval-augmented generation (RAG) applications. Isaacus’ flagship embedding model is Kanon 2 Embedder, which currently ranks as the most accurate legal AI embedder on the Massive Legal Embedding Benchmark (MLEB).
ModelArchitectureDimensionsContext windowDescription
kanon-2-embedderKanon 21,792 (default), 1,024, 768, 512, 25616,384The most accurate legal AI embedder on the Massive Legal Embedding Benchmark (MLEB).

​
Extractive question answering

Extractive question answering or answer extraction is the process of extracting answers to questions from documents. Extractive QA can be used to solve a broad range of information extraction problems from pulling out key dates from a contract to highlighting legal doctrines relied upon in a judgment. We offer two extractive QA models, tailor-made for common legal information extraction use cases: Kanon Answer Extractor and Kanon Answer Extractor Mini.
ModelArchitectureContext windowDescription
kanon-answer-extractorKanon512Our base answer extractor, designed to balance precision with throughput.
kanon-answer-extractor-miniKanon512A lighter version of the Kanon Answer Extractor, designed to maximize throughput and cost-efficiency.

​
Universal classification

Universal classification (also known as zero-shot classification) refers to the process of determining whether a given statement about a document, for example, “this is a confidentiality clause”, is supported by that document. Unlike traditional classifiers, universal classifiers don’t require training data to classify a document. In practice, universal classifiers can be used for a broad range of information retrieval and extraction tasks, from pulling out particular types of clauses from contracts all the way to scoring the relevance of legal documents to search queries. Isaacus currently offers two universal classifiers: Kanon Universal Classifier and Kanon Universal Classifier Mini.
ModelArchitectureContext windowDescription
kanon-universal-classifierKanon512Isaacus’ powerful flagship universal classification model.
kanon-universal-classifier-miniKanon512A more lightweight variant of the Kanon Universal Classifier, optimized for speed and cost-efficiency.

​
Reranking

Reranking refers to the process of scoring and sorting a set of documents based on their relevance to a particular query. Reranking is typically used in conjunction with a search engine, where the search engine retrieves a set of documents based on a user’s query, and the reranker scores and sorts those documents to present the most relevant ones to the user (or, when used as part of a retrieval-augmented generation (RAG) pipeline, to present the most relevant documents to a generative model). All of our universal classifiers also function as rerankers. They are accessible through our dedicated reranking endpoint.
⌘I