Agentic Document Extraction Drops From 135 Seconds to 8

Andrew Ng's May 2025 announcement concerns the time needed to turn a PDF into information another application can use. LandingAI's accompanying demonstration describes a document extraction approach that views a page visually rather than only reading its text, and converts text, tables, charts and form fields into Markdown that a downstream language model or retrieval app can use.

Ng walks through a report table, a traffic-accident form with multiple columns, and an invoice. The form illustrates why reading order matters: the tool infers an order across the left, middle and right columns so text stays associated with the right fields before software can use it. The post reports median processing time falling from 135 seconds to eight. That is a vendor-reported speed improvement, with no stated test-document mix. It makes the release worth evaluating for document pipelines, while leaving speed and extraction accuracy on a team's own documents to be checked.