Isaacus

In the coming weeks, Isaacus will be releasing Kanon 2 Enricher to a select group of trusted design partners via the Isaacus Beta Program. Kanon 2 Enricher is the world's first hierarchical document enrichment and graphitization model. In mere milliseconds, it can convert a lengthy, unstructured document into a rich, structured knowledge graph of sections, persons, locations, terms, citations, cross-references, dates, quotes, and more, as shown in the screenshot deconstructing the US Constitution. Although its full feature set remains under embargo, it includes hierarchical document parsing, hierarchical entity extraction, disambiguation, classification, and linking, and automatic text tagging. In total, Kanon 2 Enricher can represent 24 types of nodes, 40 types of links, and 14 types of classifications. Thanks to its unique, graph-first architecture, Kanon 2 Enricher is structurally incapable of hallucinating (though it can still misclassify). It is also extremely computationally efficient, being small enough to run locally on a consumer PC with sub-second latency. This makes it especially well-suited for applications demanding high-quality, hallucination-free, real-time, and private AI-powered document enrichment. Qualified participants in the Isaacus Beta Program will be able to experience and utilize all of the features supported by Kanon 2 Enricher before it is released to the wider public. Their feedback will be critical in helping shape our roadmap for the future of Kanon 2 Enricher and Isaacus. We will also be sharing participants' stories and use cases upon release of our products. Those interested in joining the Isaacus Beta Program are encouraged to apply here: https://isaacus.com/beta

Anyone building contract ingestion pipelines knows how difficult it is to correctly link clauses and map their dependencies. We've tackled this at Spinal using graph representations to model inter-clause relationships; but it's a constant iteration. But what if we could push that complexity down to the model level? Enter Umar Butler and the Isaacus team. We already use their Kanon 2 Embedder, and this new model looks like a significant leap forward. Excited to test it.

Here's a sneak peek into what we're working on right now at Isaacus. We’re building a new type of AI model to convert legal data into structured knowledge graphs. In other words, we’re finally making ‘law as code’ a thing. Our model, Kanon 2 Enricher, will take legal documents and, in a matter of seconds, transform them into a rich graph of clauses, terms, headings, citations, people, companies, governments, locations, dates, and more. Driving this model is the Blackstone Legal Ontology, named after Sir William Blackstone (lawyers will get it), which can currently represent 24 types of entities, 40 types of relationships, and 2,054 unique classifications, in an infinite number of permutations. There are no direct parallels to what we’re building. Kanon 2 Enricher has been designed from the ground up to output—not tokens—but an entire legal knowledge graph, and to do so over very long contexts. For fellow AI nerds out there, yes, that means that our model literally has a citation decoder head, multi-level jurisdiction classification head, clause extraction head, and so on (and, yes, that does mean we end up with >50 heads and >80 different loss terms…). Also, yes, we did manage to come up with a new way of achieving near-linear scaling on long documents. We’re aiming for release in February. By then, you’ll be able to build your own private LexisNexis without forking out billions to have your data manually labeled. In fact, Kanon 2 Enricher will offer a broader range of data types than anything you’ll find on LexisNexis or Thomson Reuters. In the meantime, enjoy this screenshot of LinkedIn's terms of service automatically annotated into the Blackstone Legal Ontology, with key entities extracted and highlighted. If you'd like to become an Isaacus Design Partner and get early insight into what we're building, reach out!

𝐈𝐬𝐚𝐚𝐜𝐮𝐬 𝐣𝐨𝐢𝐧𝐬 𝐭𝐡𝐞 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧 𝐞𝐜𝐨𝐬𝐲𝐬𝐭𝐞𝐦 🦜 We're excited to announce that LangChain now offers first-class support for Isaacus' legal AI platform! You can access our state-of-the-art legal embedding model, Kanon 2 Embedder, today via the official Isaacus LangChain integration. Thanks to Mason D. for helping make this happen. To get started, you can check out the quickstart guide on LangChain's docs here: https://lnkd.in/gfdDx4B3

Here's why fine-tuned embeddings matter so much for law. I asked the best embedding model from Google, OpenAI and Isaacus the same question: "Sally is accused of cultivating narcotic plants in her backyard. One of the elements of this charge is that “the accused intentionally cultivated or attempted to cultivate a particular substance.” To establish whether this is the case, the judge believes it would be valuable to visit Sally’s backyard and have the jury examine it for themselves. What is the name of the legal procedure whereby the court travels to a location relevant to the charge?" Open AI's text-embedding-3-large retrieved 0 documents relevant to the question, and instead was distracted by the context/additional information surrounding narcotics. Google's gemini embedding 001 did not retrieved the gold document either, nor any documents relevant to the question. Meanwhile, Kanon 2, which is legal domain specialised, had all five of its retrieved documents relate to the question. It also retrieved the gold document at k=2 and did not get distracted by mention of narcotics. What this suggests, which is obvious if you've actually played around with Kanon 2, is that this embedding model truly does have vector representations for legal concepts and procedures like a 'view'. In my observation, the performance of generic embedding models deteriorate to the level of keyword matching when dealing with specialised subject material. This is especially true for law. The embedding results will become highly biased by the words included in the query, rather than the semantic meaning of the query. I mean, it is kind of remarkable that Google and OpenAI's embeddings models retrieved documents relating to narcotics, when narcotics have nothing to do 'legal procedures' as a concept. This is despite the fact the corpus this RAG system is using has plenty of examples of other legal procedures it could retrieve instead! This is why I am working on a benchmark that will truly stress test these systems, and yes, that includes Kanon 2. If you are using these nonspecialised models in your legal RAG pipeline, I have done enough tests with these models to know just how weak they are when compared to Kanon 2. I know that sounds like shameless self-promotion, but the difference truly is night and day. And I don't think anyone has sounded the alarm yet for how powerful fine-tuning is, because besides Claude Code, there are very few large SOTA domain specialised models. More to come. #legalAI #legalembedding #legalRAG

𝘀𝗲𝗺𝗰𝗵𝘂𝗻𝗸 𝗵𝗶𝘁𝘀 𝗼𝗻𝗲 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗺𝗼𝗻𝘁𝗵𝗹𝘆 𝗱𝗼𝘄𝗻𝗹𝗼𝗮𝗱𝘀 🧩 Isaacus' semchunk library just hit one million monthly downloads today, cementing its status as the world's most popular sematic chunking algorithm, used by Microsoft's Intelligence Toolkit, Docling by IBM, and The World Bank. Leveraging complex heuristics, semchunk can break up entire books into smaller chunks in less than a second while preserving as much contextual information as possible. semchunk has now become an integral part of the retrieval-augmented generation process, enabling users to improve retrieval accuracy while saving costs and reducing hallucinations from context overflow. Isaacus remains committed to delivering superior chunking and RAG solutions to the AI community, particularly in the legal domain. We recently released the Kanon 2 Embedder embedding model, ranked first at legal information retrieval on the Massive Legal Embedding Benchmark, and are now working on Kanon 2 Enricher, an entirely new type of AI model that will perform hierarchical document segmentation, mapping divisions, articles, clauses, paragraphs, and items within legal documents to their exact character offsets; in addition to extracting and linking entities, including people, organizations, locations, key dates, citations, and more. If you're interested in working with us to improve legal AI for all, reach out! 📞 Get in touch: https://lnkd.in/gV9wR2X9 🧩 semchunk: https://lnkd.in/gXBfnHWX 📢 Announcement: https://lnkd.in/gEQ6pWQ8

𝗜𝘀𝗮𝗮𝗰𝘂𝘀 𝗷𝗼𝗶𝗻𝘀 𝘁𝗵𝗲 𝗟𝗹𝗮𝗺𝗮𝗜𝗻𝗱𝗲𝘅 𝗲𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺 🦙  Isaacus’ state-of-the-art legal AI embedding model Kanon 2 Embedder has officially landed in the LlamaIndex agentic AI framework. This integration will enable LlamaIndex users to quickly embed thousands of legal documents with Kanon 2 Embedder and then later rerank them with the Isaacus SDK without needing to reengineer existing pipelines. Kudos to Clelia Astra Bertelli for helping the Isaacus team ship this much needed integration in record time. To get started, read the Isaacus LlamaIndex integration guide on their official docs here: https://lnkd.in/gfgeEDvb

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝘁𝗵𝗲 𝗜𝘀𝗮𝗮𝗰𝘂𝘀 𝗦𝘁𝗮𝗿𝘁𝘂𝗽 𝗣𝗿𝗼𝗴𝗿𝗮𝗺 🚀 We're launching the Isaacus Startup Program, which will support legal tech startups to build new solutions with our state-of-the-art legal AI models. Participants get 𝗨𝗦$𝟱,𝟬𝟬𝟬 𝗶𝗻 𝗳𝗿𝗲𝗲 𝗔𝗣𝗜 𝗰𝗿𝗲𝗱𝗶𝘁𝘀 for their first four months of usage of Isaacus models, followed by a 𝟱𝟬% 𝗱𝗶𝘀𝗰𝗼𝘂𝗻𝘁 on model usage for the remaining seven months. Participants will also be granted 𝗼𝗻𝗲 𝘆𝗲𝗮𝗿 𝗼𝗳 𝗳𝗿𝗲𝗲 𝗽𝗿𝗶𝗼𝗿𝗶𝘁𝘆 𝘁𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 as well as access to our one year of free Isaacus priority technical support as well as access to 𝗴𝗿𝗼𝘄𝗶𝗻𝗴 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝗼𝗳 𝘃𝗲𝗻𝘁𝘂𝗿𝗲 𝗰𝗮𝗽𝗶𝘁𝗮𝗹, 𝗰𝗹𝗼𝘂𝗱 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲, 𝗮𝗻𝗱 𝗹𝗲𝗴𝗮𝗹 𝘁𝗲𝗰𝗵 𝗽𝗮𝗿𝘁𝗻𝗲𝗿𝘀 where relevant. Interested startups are encouraged to apply via our startups page. 📝 Apply https://lnkd.in/g4UDvEcY 📢 Announcement https://lnkd.in/gnG46rzM

𝗜𝘀𝗮𝗮𝗰𝘂𝘀 𝗷𝗼𝗶𝗻𝘀 𝘁𝗵𝗲 𝗛𝗮𝘆𝘀𝘁𝗮𝗰𝗸 𝗲𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺! As part of our mission to make the world’s best legal models available to AI engineers wherever they need them, we’ve expanded our model offerings to deepset's Haystack framework. From now, our state-of-the-art legal embedding model, 𝗞𝗮𝗻𝗼𝗻 𝟮 𝗘𝗺𝗯𝗲𝗱𝗱𝗲𝗿, is available via the Haystack embedder pipeline. Our rerankers and classifiers will follow in the coming months, with additional, as-yet-unannounced solutions also planned for Haystack. A big thank-you to the engineers at deepset for guiding us through the integration and making it smooth and fast. We’re excited to see what you build with our integrations. 🗒️Isaacus x Haystack integration docs: https://lnkd.in/gKNZbzV9 💻Isaacus x Haystack Github repo: https://lnkd.in/gkD8V9Ak 🚀Isaacus quickstart guide: https://lnkd.in/gyTPKj3Z