Harvey redesigns Review Table loading for larger legal document sets
Harvey said it redesigned Review Table loading so large legal document reviews open faster and remain responsive as row counts and column counts grow. The engineering update matters because the company described Review Table as a collaborative system for analyzing thousands of documents, with each row tied to a file and each cell backed by model reasoning and citations.
Evidence and context
In a blog post published on 22 October 2026, Harvey said Review Table is its largest and most exhaustive inference engine, used to standardize answers in a structured table that lawyers can sort, filter, verify, lock and comment on. Harvey said the feature is intended to become a reusable source of truth for later memos or presentations, but the browser-based loading approach began to break down as teams queried hundreds of columns across thousands of files.
Harvey said the old design performed all table operations on the client side, loading the entire Review Table into the browser and handling sorting, filtering and aggregation in memory. That approach worked for smaller tables, but the company said tables could lag, become unresponsive or fail to load entirely once they exceeded thousands of rows.
To address that limit, Harvey said it moved client-side operations to the server side and adopted a CQRS-style split between Postgres, which remains the source of truth, and Turbopuffer, which provides the read interface. Harvey said Turbopuffer fit the system because it runs on top of Azure blobs and supports vector search, full-text search and filtering, but it also required new abstractions because it does not support pagination out of the box.
The company said it implemented ID-based pagination: the backend returns ordered row IDs, and the client requests data for arbitrary lists of row IDs as users scroll. Harvey said this allowed variable-width page loads while keeping overhead lower than a cursor or limit/offset scheme would have. The company also said it had to support data-type-aware sorting and filtering by dual-projecting strict types and string representations for non-string cells such as dates, currencies and classifications.
Harvey said it also built safeguards for schema drift and schema evolution. If a Turbopuffer write fails or drift is detected, reads fall back to Postgres while a backfill job repairs the broken namespace. The post said the system writes table skeletons and cell data across Postgres, Azure blob storage and Turbopuffer, with batching on the Turbopuffer side.
On the frontend, Harvey said it now loads Review Tables in phases: first a skeleton with table status, columns and file metadata; then the values in the sort column; and only after that the rows for the current page. The company said the table outline is rendered before the data cells arrive, with a loading animation used in the interim. As users scroll or change sort and filter parameters, the system fetches additional pages and caps the number of rows held in memory.
Harvey said the redesign produced immediate latency improvements across all table sizes, with larger tables seeing bigger gains. The company reported speedups of 1.2x for tables under 1,000 cells, 2.2x for 1,000 to 10,000 cells, 5.8x for 10,000 to 100,000 cells and 4.2x for 100,000 to 500,000 cells. Harvey said tables that previously failed to load now open successfully.
The timing follows a run of recent Harvey announcements, including a 21 October 2026 newsroom post saying CVS Health deployed Harvey across its corporate affairs functions, and earlier September 2026 disclosures about Guardrails AI joining Harvey and a $550 million financing at a $15.5 billion valuation. The company’s latest engineering post does not provide a next product date.