TechPrufer
Get started
Harvey
Harvey logoHarvey · Feature trail

Vault Upload Performance and Reliability

7 posts · Nov 2024 – Sep 2026

What this trail covers

Harvey's Vault system allows deal teams to process large volumes of documents, perform bulk extractions of key provisions (e.g., change-of-control, assignment restrictions, indemnity caps), and surface these results in structured review tables linked to source documents. This significantly reduces review time and enables reviewers to move directly to findings. The system is designed for enterprise readiness, with integrations and a focus on traceability and auditability. The capability is further enhanced with increased file limits per project (10,000 documents), one-click workflows for common document types with high recall scores (e.g., Merger Agreements at 99.66%), customizable review tables with proactive data type suggestions, and integration into Assistant for seamless querying of Vault documents. Firms can also create and share custom extraction workflows. The system continuously works with the latest embedding and storage systems to improve document processing time and limits.

All Harvey trails
  1. Agentic Vault Search and Organization in HarveyThis post details the technical implementation of Agentic Vault Search and Organization. Agentic Vault Search leverages signals like folder location, naming conventions, and uploader identity to locate documents, moving beyond simple keyword matching. Agentic Vault Organization allows users to specify desired folder structures or renaming conventions in natural language, with the system proposing and confirming changes before execution. Key technical considerations include agent permissions, audit trail logging, and ensuring content integrity by not editing file contents.
  2. What Goes Into an Accurate and Trusted Legal Deal SummaryThis post details how Harvey builds trustworthy legal deal summaries by emphasizing source-based output with direct links to exact clauses, a visible production record for auditability, clear handling of ambiguous input, a required human review step, and consistent application of standard positions. It highlights the use of Harvey's Vault capability for extracting key terms from transaction documents and assembling summaries, with a specific mention of Bridgewater's use case. The post also frames deal summary generation as an AI agent workflow, where a discrete workflow agent extracts terms and generates a first-pass, fully cited summary, which is then reviewed by a lawyer. The emphasis is on grounding the summary in source documents to ensure accuracy and trust, making the lawyer's review fast and certain.
  3. How Harvey’s AI is Changing Data Room Due Diligence in M&AThis post details how Harvey's Vault system has evolved to address the challenges of data room due diligence in M&A. It explains how the system enables bulk extraction of key provisions across entire contract sets, organizing findings by workstream and exception category rather than by individual documents. The contribution highlights the shift from document-by-document review to exception review, allowing lawyers to focus on evaluating risk. It emphasizes the defensibility of findings through source-cited, checkable results and a clear audit trail, aligning with professional judgment and ABA ethics guidance. The post also outlines a day-by-day workflow from data room opening to bid deadline, showcasing how AI-driven diligence accelerates the process and facilitates collaboration through shared vaults.
  4. Making Vault Uploads Faster and More Reliable | HarveyThis post details the migration of Vault's file upload architecture from server-proxied uploads to a presigned URL model. This involved splitting the control plane (backend) from the data plane (storage). The frontend was re-architected into a three-stage concurrent pipeline to manage uploads directly to Azure Blob Storage. The backend's role shifted to authorization, validation, and finalization, with optimizations for batched operations to maintain scalability. Production safety was enhanced by controlling browser/network concurrency and preserving backend correctness through tight transaction boundaries and idempotent operations.
  5. Building a New File Ingestion System to Scale Firm KnowledgeIntroduced a new asynchronous file ingestion system using Temporal for orchestration, enabling folder uploads and one-way sync from DMS. Implemented a granular, activity-based workflow for resilience against rate limits and transient failures. Developed an in-house rate limiter on Redis to manage external API quotas. Enhanced the sync algorithm with hash-based content change detection and careful ordering of operations for correctness. Optimized performance through aggressive batching, connection pooling, parallelization, and careful task queue design.
  7. Scaling Harvey’s Document Systems: Vault File Upload and ManagementThis post details significant architectural improvements to Harvey's Vault system, specifically focusing on scaling file upload and management for large document sets. Key technical contributions include: isolating upload operations into dedicated web workers, implementing separate state management for temporary vs. processing errors, building a robust file reconciliation system, and enabling intelligent parallel batching. The frontend dynamically adjusts parallelism based on file metadata and system load, while the backend consolidates queries into efficient bulk operations. For interface performance, the system migrated compute-intensive operations (sorting, filtering, aggregation) to the server-side, reducing browser memory consumption by 90% and improving initial load time by 80%. This involved developing server-driven search and file lists, cross-page file selection with lazy parent-mapping, event-driven context architecture, new bulk operation APIs, smart caching with predictive prefetching, and adaptive progress indicators.
  8. Introducing the Next Version of VaultThis post details the upgrade of Harvey's Vault system, increasing the file limit per project from 1,000 to 10,000 documents. It introduces one-click workflows for common document types (e.g., Merger Agreements, Stock Purchase Agreements, Leases) with specific recall scores and field counts, engineered by legal and AI researchers. The review tables are redesigned for flexibility, allowing individual column configuration with data type selection (dates, currency, verbatim text) and proactive suggestions. Vault is now integrated into Assistant for querying without page switching, and both Assistant threads and Vault projects can be shared. The system is continuously working with the latest embedding and storage systems to improve document processing time and limits.