Release Notes

Stay up-to-date with the latest improvements, new features, and bug fixes for Kensho Extract.

August 2026

Documents up to 4,000 pages. Raised from 3,500.
Self-deployed container: OCR enabled by default. No longer requires opt-in. Disable with ENABLE_DOCUMENT_OCR=false if needed.

Extractions are no longer lost during scale-down. Workers now finish the extraction they're running before shutting down.
Bug fixes and stability improvements.

Multi-page tables merge into one result. Set merge_multi_page_tables on the v3 extract request and a table that continues across a page break comes back as a single table instead of one per page, with no repeated header rows. Also available in the self-deployed container.

Documents up to 3,500 pages. Raised from 3,000.
Self-deployed container: layout analysis matches the hosted service. The container was pinned to an older document-layout model, so the same document could come back structured differently depending on where you ran it. Both now use the same version.

Text and table coordinates are now correct on rotated and landscape pages. Positions on those pages were previously reported against the unrotated orientation.
Password-protected PDFs now return a bad-document error naming the problem, instead of a generic service error that looked like an outage on our side.

Force the OCR language. The v3 extract request takes an optional language, which is passed to OCR instead of letting it auto-detect. Useful when auto-detection picks the wrong language on short or mixed-language pages.
Table row header hierarchy. Set extract_table_row_header_hierarchy and nested row headers come back with their parent-child structure, so indented row labels keep their meaning instead of arriving as a flat list.
Self-deployed container: HTTP interface. The container now runs an HTTP server, so you no longer need Kafka to send it work. It's close to a drop-in replacement for the queue-based container, and the release ships a migration guide.
Self-deployed container: built-in OCR. The container does its own OCR, so scanned documents no longer require a connection to the hosted OCR service.

Page-by-page OCR is now on for everyone. Faster and more accurate on scanned documents. Rolled out gradually through May and June, finishing at 100% in June.
Documents up to 3,000 pages. Raised from the previous limit.

Pages with thousands of tiny text segments no longer stall. Text ordering now falls back to a faster method past a size threshold.
Runaway extractions now time out. Time limits weren't being enforced, so a single stuck document could run for hours and block faster requests queued behind it.
No more duplicate processing. Fixed acknowledgement handling that could cause the same request to be processed more than once.
Stability improvements.
OCR jobs are no longer dropped when infrastructure fails.

Tag uploaded files. POST /v3/extractions/upload-url accepts file_tags, and the tags are written onto the stored objects for the request and its output.

Fixed an inverted y-axis bug that put table and text positions upside down in some documents.

Supporting relations between segments. Output can now include which segments support which, for example a footnote and the row it annotates.
Self-deployed container: figure OCR. Text inside charts and figures is now extracted in the container.

Fixed incorrect hierarchy assignment when merging multi-part documents, which produced wrongly nested sections.