Harvey AI Builds Enterprise File Ingestion System for Legal Firms
Jessie A Ellis Feb 12, 2026 05:53
Legal AI startup Harvey unveils high-throughput file ingestion system capable of processing hundreds of thousands of documents from enterprise DMS platforms.
Harvey AI has rolled out a new file ingestion architecture designed to handle hundreds of thousands of legal documents from enterprise document management systems, addressing a critical bottleneck in how law firms feed institutional knowledge into AI tools.
The system targets a fundamental problem: large law firms sit on millions of documents containing deal structures, motion templates, and negotiation playbooks scattered across platforms like iManage, SharePoint, and Google Drive. Getting that context into AI systems—and keeping it fresh—has been a manual nightmare.
What Changed
Harvey's previous approach relied on synchronous file processing with manual uploads. Users had to select individual files rather than folders, and documents went stale whenever someone updated the source. The new system introduces two core features: one-click folder uploads that preserve entire hierarchies with metadata, and continuous one-way sync that automatically detects and pulls changes from connected DMS platforms.
The technical backbone uses Temporal for workflow orchestration—a choice driven by the unpredictable nature of enterprise file operations. Traffic spikes, external rate limits, and transient failures are constants when crawling millions of documents across distributed systems.
The Engineering Tradeoffs
Harvey's team made each API request a separate Temporal activity during folder crawling. This granularity means hitting a rate limit on page 47 of a 200-page folder listing triggers a retry for just that request, preserving progress on the previous 46 pages. File downloads follow the same isolation pattern—one file failing doesn't tank the batch.
Rate limiting turned out to be the hidden complexity. Each integration partner enforces limits differently: by request count, payload size, or both; scoped per-user or per-organization; sometimes changing dynamically based on tier. Harvey built custom rate limiters on Redis that track every external request and proactively throttle before hitting limits—critical for ensuring background ingestion jobs don't degrade real-time user features.
Market Context
The timing aligns with broader enterprise AI infrastructure moves. Matia raised $21 million on February 10 to consolidate data management including ingestion systems. EXL secured 10 new U.S. patents on February 9 covering multimodal data ingestion and knowledge graph creation. The data integration market is projected to hit $17.1 billion in 2025, with ETL tools reaching $7.63 billion by 2026.
For legal tech specifically, the play is straightforward: AI output quality correlates directly with the quality and recency of context fed into it. Answers grounded in a firm's actual precedents beat generic responses. The question is whether Harvey can maintain that context pipeline at enterprise scale without becoming a maintenance burden.
What's Next
Harvey says the architecture extends beyond files to client matter metadata, ethical walls, emails, and billing entries. The same infrastructure could power real-time AI agents that search DMS platforms directly without pre-ingestion. Adding new DMS integrations now takes days instead of weeks—the core orchestration, rate limiting, and processing pipeline are shared.
The features are currently in early access with general availability planned soon. For firms already drowning in document management overhead, the pitch is simple: stop manually maintaining AI context and let the system handle it.
Image source: Shutterstock.jpg)