Small models as paralges: lexisnexis distill models for building AI assistants

As a legal research company Lexisnexis He created his Ai Assistant Protegé and wanted to find out how best to use his know -how without using a large model.

Protégé aims to help lawyers, employees and paralregals, to write and prove legal documents and to ensure that everything they cite Complaints and letters are correct. However, Lexisnexis did not want a general legal assistant for legal readiness. They wanted to build one who learns a company’s workflow and is more customizable.

Lexisnexis saw the opportunity to bring the power of large -scale language models (LLMS) from anthropic and mistal and to find the best models that answer the best to users.

“We use the best model for the specific application as part of our multi-model approach. We use the model that delivers the best result with the fastest response time,” said Riehl. “For some applications, this is a small voice model like Mistral or we carry out distillation to improve performance and reduce costs.”

While LLMS still has a value when building AI applications, some organizations in small voice models (SLMS) or distilling LLMS turn to small versions of the same model.

Distillation, where a LLM “teaches” a smaller model, has become a popular method For many organizations.

Small models are often best suited for apps such as chatbots or simple code design, which Lexisnexis wanted to use for Protegé.

This is not the first time that Lexisnexis AI applications have created even before the Legal Research Hub Lexisnexis + AI started in July 2024.

“We have used a lot of AI in the past, which was more on the processing of natural language, a little deep learning and mechanical learning,” said Riehl. “That really changed in November 2022 when chatt started, because before that many of the AI ​​functions were behind the scenes. But as soon as Chatgpt came out, the generative skills were very fascinating for us.”

Small, finely coordinated models and model routing

According to Riehl, Lexisnexis uses various models from most main model providers when building its AI platforms. Lexisnexis + AI used Claude models of Anthropic, Openai -Gpt models and a model from Mistral.

This multimodal approach has contributed to breaking up the individual tasks that users wanted to carry out on the platform. To do this, Lexisnexis had to architect its platform switch between models.

“We would break up every task into individual components and then identify the best major language model to support this component. An example of this is that we will use Mistral to assess the query in which the user was entered,” said Riehl.

For Protegé, the company wanted faster response times and models that are better coordinated for legal applications. So it turned to what Riehl described “fine” versions of models, essentially smaller weight versions of LLMS or distilled models.

“You do not need a GPT-4O to assess the assessment of a query. That is why we use it for more sophisticated work and switch models,” he said.

If a user asks Protegé to a certain case, the first model that IT Pings has pings is a finely coordinated Mistral-Ratz for assessing the query and then determining what the purpose and the intention of this query is “before it is converted to the model, which is best to do the task. Generated that summarizes the results.

At the moment, Lexisnexis mainly relies on a finely coordinated Mistral model, although Riehl said that it used a finely coordinated version of Claude, “when it came out for the first time; we do not use it in the product today, but in other ways.” Lexisnexis is also interested in using other Openai models Reinforcement of fine -tuning skills last year. Lexisnexis evaluates the justification models from Openai, including O3 for its platforms.

Riehl added that it could also examine the use of Gemini models from Google.

Lexisnexis supports all AI platforms with their own knowledge graph to perform the functions of the AUGMENDEDEDEDEDE (RAG) call, especially since Protegé could later help start agent processes.

The Ki Legal Suite

Lexisnexis tested the opportunity to work in the legal industry before the generative AI arises. 2017 the Company tested an AI assistant This would compete with IBMS Watson companies from IBM in Ross and Protegé, which are on the company’s Lexisnexis + AI platform, which brings together the AI ​​services of Lexisnexis.

Protégé helps law firms with tasks that tend to do paralegals or employees. It helps to write legal investigations and complaints based on documents and data from the companies, suggest the legal workflow next steps, suggest new requests for refining search processes, draft questions for deposits and discoveries, generate quotes for accuracy, timetables and of course summarize complex legal documents.

“We see Protegé as the original step in personalization and agent skills,” said Riehl. “Think of the different types of lawyers: M&A, legal dispute, real estate. It will continue to be more personalized due to the specific task you do.

Protégé is now competing against other legal research and technology platforms. Thomson Reuters have adapted Openas O1-Mini model for his COCOUSEL Legal Assistant. Harvey, what collected $ 300 million Investors such as lexisnexis also has a legal AI assistant.