Datalab

Datalab grew out of the open-source document AI community behind Marker, Surya, Chandra and Indra — models with tens of thousands of GitHub stars between them — and packages that research into a managed, production-grade API. The platform organises its capabilities as "processors": Convert handles PDF-to-Markdown/HTML/JSON with optional word-level bounding boxes and chart data, Extract pulls user-defined fields into JSON with per-field citations and confidence, Segment splits combined files on natural document boundaries, and Eval compares parsing modes against your own documents.

Because the team owns its models, it can offer unusually flexible deployment: a managed SOC 2 cloud with a free tier, a VPC deployment inside your AWS, GCP or Azure account, or a fully air-gapped on-premises install with no internet access. Public benchmark work — notably the open OmniParseBench, with over sixteen thousand pass/fail tests — is used to argue that its accurate mode leads competing parser APIs, especially on tables and multi-column layouts.

Typical buyers are frontier AI research labs building training corpora, plus regulated verticals such as financial services, healthcare, insurance and legal, where traceability matters as much as raw accuracy. Batch processing is pitched at very large volumes — the company describes managed batch runs handling a billion-plus pages a day.