Harvey rebuilds large legal Review Tables around server-side loading

An engineering update moves sorting and filtering away from the browser while retaining document-backed answers and collaborative review.

Harvey has redesigned the loading architecture behind Review Tables, its interface for comparing answers across large document collections. The company's [22 October engineering account](https://www.harvey.ai/blog/making-review-tables-faster-and-more-reliable) describes a move from loading an entire table into browser memory to fetching the rows needed for the current view.

The tables place documents in rows and natural-language questions in columns, with cited answers in individual cells. Teams can assign rows, comment, verify answers and lock them. The new design keeps Postgres as the source of truth and adds a Turbopuffer read layer, while retaining separate Azure storage for sensitive customer content.

The expensive part happens after the model answers

A table can become difficult to use even when individual AI answers arrive successfully. Thousands of documents multiplied by many questions create a large working dataset. Sending that entire collection to the browser means the user's device must hold the data, sort it, filter it and keep the interface responsive.

This is a different performance problem from model inference. Faster token generation cannot remove a browser's memory limit. The application needs a way to display a useful portion of the review while preserving the user's position in the larger collection.

Harvey's approach returns an ordered set of row identifiers, then loads selected rows. Sorting or filtering changes that ordering and triggers new page requests. A preliminary table outline provides structure while cell content arrives.

Large reviews need stable navigation

For a professional reviewing agreements, the scrollbar is part of the task. A user may jump to a folder, compare the final rows with the first ones, or change a date filter while preserving a selection. Pagination that only supports moving forward one page at a time would make some of those interactions awkward.

Separating row order from row content provides more flexibility. The interface can know which documents belong in a particular position without keeping every answer in memory. It also makes a loading state more informative: the professional can see the table's shape and the files being reviewed before all answers appear.

This pattern is useful beyond legal software. AI-generated comparison tables in procurement, insurance and research can face the same gap between a completed inference job and an interface that remains practical at scale.

Typed answers and a recoverable read path

The engineering account also describes storing typed values alongside string representations, so dates and currencies can be sorted by their meaning. If the secondary read layer falls out of sync, requests return to the original data path while a backfill repairs the affected namespace.

These are valuable details because an AI answer is not always a clean database field. A date might appear in a sentence; a financial value might include an explanation. A user expects chronological or numerical sorting, while the application still needs the original answer for review.

The update shows how much of professional AI depends on ordinary application engineering. Document analysis becomes useful when people can navigate, compare, discuss and reuse its results. Harvey's redesign concentrates on that part of the workflow: keeping the review surface usable as the amount of generated work grows.

Sources

- [How We Made Review Tables Faster and More Reliable](https://www.harvey.ai/blog/making-review-tables-faster-and-more-reliable) — Harvey