# Harvey — How we Built Image Understanding for Legal Documents

- Company: Harvey (harvey.ai)
- Announced: 2026-05-06T18:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/building-image-understanding-for-legal-documents
- Record: https://forck.live/items/17571-how-we-built-image-understanding-for-legal-documents
- Subject: Harvey

Harvey introduced an on-demand visual analysis system for its legal AI assistant. The system renders document pages at high resolution and uses a vision model to read charts, diagrams, and complex layouts only when a user's query requires visual understanding. It first searches document text to identify candidate pages, and gates image analysis behind a text search step to avoid unnecessary processing.

## Evidence

Verbatim from https://www.harvey.ai/blog/building-image-understanding-for-legal-documents:

> Rather than having Harvey describe every image at indexing time (a process that is expensive, lossy, and often wrong without context), we built an on-demand visual analysis tool that the agent invokes only when it needs to.

---

Record: https://forck.live/items/17571-how-we-built-image-understanding-for-legal-documents
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md