Introducing the Next Version of Assistant

Harvey’s most used product is Assistant – a professional AI tool that drafts documents, analyzes information, and answers questions across hundreds of files. Every day, tens of thousands of legal, tax, and finance professionals rely on Assistant for a wide range of tasks, from transactional work to litigation to audits and beyond (more info here).

Today, we are thrilled to release a revamped version of Assistant, rolling out to all users over the next few weeks.

This updated version brings three major enhancements: (1) specialized modes tailored to different use-cases, (2) iterative functionality for refining and expanding on initial responses, and (3) faster processing, higher-quality outputs, and enhanced document processing capabilities.

1. Mode Choice

Assistant is now divided into multiple modes, starting with Assist and Draft.

Assist Mode is optimized for summaries, analyses, and searches. Example: Identify inconsistencies across these depositions. Provide in a table format.
Draft Mode is purpose-built for generating and revising detailed long-form content like briefs, contracts, contractual provisions and more with a legal/professional style. Example: Based on these precedents, draft an NDA for my client who is a private equity firm based in Delaware, for a potential acquisition of a target company.

Our research indicates that using different algorithms and underlying models for specific tasks results in stronger outputs.

2. Iteration on answers

The most common request from Harvey users has been to incorporate iterative functionality to expand on Harvey’s first answer. We’re excited to launch two ways to do this – users can engage in follow-up questions in Assist Mode or revisions in Draft Mode. In each instance, Harvey will produce tailored content based on the document and previous answers.

Assist Mode: Users can ask free-form follow-up questions (e.g., make this more concise) or utilize AI-generated suggestions based on the most recent output.

Draft Mode: Users can prompt Harvey to revise the output or highlight specific text for focused revisions. "Show Edits" allows users to track changes between revisions.

Our users often seek detailed, in-depth answers for large professional documents. This requires an AI system that preserves memory for long prompts, page-long outputs, and large documents with dense text, citations, tables, and graphics. Our research team has been focused on ensuring that we can build this system without sacrificing quality, speed, or security. This is the result of our significant advancements in agents, retrieval techniques, and evaluation methods, all powered by the latest generation of foundation models. We’re excited for everyone to try it.

3. Performance Improvements

The new version of Assistant includes many overarching performance improvements to our document understanding, sourcing, and reasoning systems. The most visible to users are:

Accuracy: Based on our internal evaluations, we’ve reduced hallucinations by 60%
Sourcing: Based on our internal evaluations, we’ve improved the accuracy of cited sources by 23%
Speed: Based on our internal tests, we’ve improved time to first word by 80%
Excel Support: Better parsing and citations for xlsx/xls spreadsheets with multiple sheets, tables, and numbers.
Language Support: higher fluency across a broader set of languages
Prompt Understanding: Improved handling of prompts containing multiple requests

Next Steps

We are committed to rapidly improving the Harvey platform to keep up with the pace of AI technology. This update will be rolling out over the next few weeks.

We also remain committed to improving transparency and helping clients better understand our product – over the next few months, we will continue sharing more insights into our product development, AI advancements, and evaluation processes to measure performance. We welcome your feedback on how your firm is using or planning to use Harvey, as we work to support our clients in adopting generative AI.