At Harvey, we aim to simplify and improve the knowledge systems that legal, finance, and tax professionals rely on every day. Vault is a collaborative workspace for large-scale document review, analysis, and synthesis. It helps professionals handle complex matters involving tens of thousands of documents in minutes, not hours or days. 

Today, we’re excited to launch an upgraded version of Vault, with a focus on speed, usability, and precision.

First, we’ve increased the file limit per project from 1,000 to 10,000 documents. This expansion supports the management and analysis of larger datasets within a single, organized environment—accommodating the demands of complex matters without compromising performance. We are continuously working with the latest embedding and storage systems to improve document processing time and limits.

A major use case we’ve seen with Vault is extracting meaningful insights from thousands of documents to support complex projects under tight deadlines. However, we found that the complexity of prompt engineering and other analysis often detracted from Vault’s efficacy for lawyers. With the introduction of one-click workflows for common document types, Harvey steps in as an intelligent partner, absorbing the burden of intricate prompt engineering and computational heavy lifting that these tasks typically demand.

These workflows are specifically engineered by our research team of former BigLaw attorneys and AI researchers. While we started with SPA Deal Points, our workflows now cover a broad range of document types across transactional, regulatory and litigation matters handled by in-house or outside counsel. With a single click, users can extract 25+ data points per document. Once finished, they can further tailor Vault’s analysis by adding or removing terms.

We evaluated our workflows on test sets consisting of representative documents. For evaluation, we compute recall compared to ground truth extractions verified by attorneys. Optimizing for recall ensures that Vault is surfacing all relevant data points, including ambiguities and judgment calls, to users for their review and analysis. Overall recall for each workflow is computed as an average over the recall score for each field.

Over the coming months, we’ll be adding new workflows tailored to additional document types. In parallel, we’re enabling firms to create and share their own custom extraction workflows, allowing them to infuse their deep expertise directly into Vault. Embedding this combined domain expertise into the platform allows all users to seamlessly access exceptional AI, spending less time configuring prompts for key use cases and more time focused on delivering value to their clients.

While workflows automate the execution of common use cases, we recognize that users often require flexibility to perform custom analyses. As a result, Vault’s review tables have been redesigned for flexibility. Users can now configure each column individually, selecting data types—like dates, currency, or verbatim text—that fit each analyses’ unique requirements. Harvey will even proactively suggest types based on the input of the column. Review tables then extract an answer that matches the column’s data type, allowing teams to tailor each review output to their specifications for maximum clarity and relevance.

For example, the “Verbatim” column type extracts a direct quote from the document, preserving its exact wording. This is particularly useful when users reviewing a contract need to pull specific clauses word-for-word.

Additionally, Vault is now integrated into Assistant, allowing users to easily query Vault documents where they already spend the majority of their time in Harvey – without switching pages. Both Assistant threads and Vault projects can be shared, enabling colleagues to collaborate on projects and exchange knowledge and insights.

This upgrade lays the foundation for ongoing improvements. As Harvey evolves, we will expand the capabilities of both Vault and Assistant, integrating more sophisticated workflows that address a broader spectrum of knowledge work tasks. We welcome feedback as we continue to help our clients achieve mastery in their domains faster than ever before.