recently,HarveyCompany andOpenAIAnnouncing a partnership for thelawProfessionals have built a custom-trained case law model. This AI system is not only capable of complex reasoning, but is also able to handle a wide range of legal domain knowledge and has the ability to go beyond a single model call.
It is capable of drafting legal documents, answering questions about complex litigation scenarios, and even identifying significant discrepancies between hundreds of contracts. This innovative initiative aims to improve the efficiency and accuracy of legal work and provide strong technical support to legal professionals.
Harvey was co-founded by Winston Weinberg, an attorney with a background in antitrust and securities litigation, and Gabe Pereira, an artificial intelligence researcher. They recognized the tremendous potential of using large-scale language models (LLMs) to synthesize information and present it to attorneys for review.
For the case law study, the Harvey team's goal was to create an experience where users could copy and paste client questions directly into the model, which would provide thorough answers and cite all sources.
To accomplish this, the Harvey team first tried obvious techniques such as fine-tuning the base model through public APIs and building a Retrieval Augmented Generation (RAG) system. However, they quickly realized that these techniques would not meet the needs of complex, open-ended use cases.
As a result, Harvey decided to work with OpenAI to build a custom training model in order to inject new knowledge and ways of reasoning about that knowledge into the base model. They started with Delaware case law and eventually expanded to case law across the United States, adding the equivalent of 10 billion tokens of data to the model.
Over the past year, Harvey has become the trusted generative AI platform for legal, tax and financial professionals. The company has grown its team to over 100 people and has grown its 2023 revenues by more than 10x. Harvey also recently secured $80 million in Series B funding from investors including EladGil, KleinerPerkins, OpenAI and Sequoia, valuing the company at $750 million.
Harvey's case law model has several functional features. It is capable of handling tasks that require complex reasoning, which is particularly important for legal professionals. The model has been custom-trained to gain knowledge covering a wide range of legal domains and is able to understand and handle a wide range of law-related queries and tasks.
In addition, the model can help legal professionals draft and review legal documents, significantly improving efficiency. When it comes to analyzing complex litigation scenarios, models can provide in-depth analysis and solutions. Customized models provide more accurate and relevant legal information and solutions than traditional models and ensure that every sentence has a clear source citation.
To test the performance of the case law model, Harvey worked with tenmaximumof law firms collaborated on an evaluation. They provided attorneys with a side-by-side comparison of customized case law model outputs with GPT-4 outputs.
The results showed that lawyers preferred the output of the customized case law model in the context of 97%. This strong preference was mainly due to the fact that the customized model provided longer and more complete answers, delved more deeply into the details of the issue, and covered more relevant case law.
Harvey's next focus is to explore how multiple model calls can be combined into a single work output to streamline the user experience. Their vision is to be a supportive member of the team, helping assistants with complex but routine tasks, allowing professionals to focus their time on client interactions. In this way, Harvey has great potential for growth not only in the legal field, but in all areas of professional services.