When authorized analysis firm LexisNexis created its AI assistant Protégé, it needed to determine one of the simplest ways to leverage its experience with out deploying a big mannequin.
Protégé goals to assist legal professionals, associates and paralegals write and proof authorized paperwork and be certain that something they cite in complaints and briefs is correct. Nonetheless, LexisNexis didn’t need a common authorized AI assistant; they needed to construct one which learns a agency’s workflow and is extra customizable.
LexisNexis noticed the chance to carry the facility of enormous language fashions (LLMs) from Anthropic and Mistral and discover the very best fashions that reply person questions the very best, Jeff Riehl, CTO of LexisNexis Authorized and Skilled, informed VentureBeat.
“We use the very best mannequin for the precise use case as a part of our multi-model strategy. We use the mannequin that gives the very best consequence with the quickest response time,” Riehl stated. “For some use instances, that will probably be a small language mannequin like Mistral or we carry out distillation to enhance efficiency and cut back value.”
Whereas LLMs nonetheless present worth in constructing AI functions, some organizations flip to utilizing small language fashions (SLMs) or distilling LLMs to turn out to be small variations of the identical mannequin.
Distillation, the place an LLM “teaches” a smaller mannequin, has turn out to be a preferred methodology for a lot of organizations.
Small fashions typically work finest for apps like chatbots or easy code completion, which is what LexisNexis needed to make use of for Protégé.
This isn’t the primary time LexisNexis constructed AI functions, even earlier than launching its authorized analysis hub LexisNexis + AI in July 2024.
“We’ve got used a number of AI up to now, which was extra round pure language processing, some deep studying and machine studying,” Riehl stated. “That basically modified in November 2022 when ChatGPT was launched, as a result of previous to that, a number of the AI capabilities had been type of behind the scenes. However as soon as ChatGPT got here out, the generative capabilities, the conversational capabilities of it was very, very intriguing to us.”
Small, fine-tuned fashions and mannequin routing
Riehl stated LexisNexis makes use of totally different fashions from a lot of the main mannequin suppliers when constructing its AI platforms. LexisNexis + AI used Claude fashions from Anthropic, OpenAI’s GPT fashions and a mannequin from Mistral.
This multimodal strategy helped break down every job customers needed to carry out on the platform. To do that, LexisNexis needed to architect its platform to change between fashions.
“We’d break down no matter job was being carried out into particular person elements, after which we might establish the very best massive language mannequin to assist that part. One instance of that’s we are going to use Mistral to evaluate the question that the person entered in,” Riehl stated.
For Protégé, the corporate needed quicker response occasions and fashions extra fine-tuned for authorized use instances. So it turned to what Riehl calls “fine-tuned” variations of fashions, primarily smaller weight variations of LLMs or distilled fashions.
“You don’t want GPT-4o to do the evaluation of a question, so we use it for extra subtle work, and we swap fashions out,” he stated.
When a person asks Protégé a query a couple of particular case, the primary mannequin it pings is a fine-tuned Mistral “for assessing the question, then figuring out what the aim and intent of that question is” earlier than switching to the mannequin finest suited to finish the duty. Riehl stated the following mannequin might be an LLM that generates new queries for the search engine or one other mannequin that summarizes outcomes.
Proper now, LexisNexis largely depends on a fine-tuned Mistral mannequin although Riehl stated it used a fine-tuned model of Claude “when it first got here out; we’re not utilizing it within the product right now however in different methods.” LexisNexis can be fascinated by utilizing different OpenAI fashions particularly for the reason that firm got here out with new reinforcement fine-tuning capabilities final 12 months. LexisNexis is within the strategy of evaluating OpenAI’s reasoning fashions together with o3 for its platforms.
Riehl added that it might additionally take a look at utilizing Gemini fashions from Google.
LexisNexis backs all of its AI platforms with its personal data graph to carry out retrieval augmented technology (RAG) capabilities, particularly as Protégé might assist launch agentic processes later.
The AI authorized suite
Even earlier than the arrival of generative AI, LexisNexis examined the potential of placing chatbots to work within the authorized trade. In 2017, the corporate examined an AI assistant that might compete with IBM’s Watson-powered Ross and Protégé sits within the firm’s LexisNexis + AI platform, which brings collectively the AI companies of LexisNexis.
Protégé helps legislation corporations with duties that paralegals or associates are likely to do. It helps write authorized briefs and complaints which can be grounded in corporations’ paperwork and information, recommend authorized workflow subsequent steps, recommend new prompts to refine searches, draft questions for depositions and discovery, hyperlink quotes in filings for accuracy, generate timelines and, in fact, summarize complicated authorized paperwork.
“We see Protégé because the preliminary step in personalization and agentic capabilities,” Riehl stated. “Take into consideration the various kinds of legal professionals: M&A, litigators, actual property. It’s going to proceed to get increasingly more personalised based mostly on the precise job you do. Our imaginative and prescient is that each authorized skilled can have a private assistant to assist them do their job based mostly on what they do, not what different legal professionals do.”
Protégé now competes towards different authorized analysis and expertise platforms. Thomson Reuters custom-made OpenAI’s o1-mini-model for its CoCounsel authorized assistant. Harvey, which raised $300 million from buyers together with LexisNexis, additionally has a authorized AI assistant.
