Category: Data & documents

Rows and columns on one side, emails, PDFs and recordings on the other, JSON somewhere in the middle. How the 3 types differ, where each one lives, and why LLMs made converting between them cheap.

Emails, PDFs, scans, chats and call recordings hold most of what a company knows, with no fields to query. What unstructured data is, how much of your data it probably is, and the 2026 toolkit for reading it.

Multimodal models read neat modern handwriting with under 2% character errors in published tests. Messy, historical and non-English pages are another story. What the benchmarks say, which tools to shortlist, and how to set up the human check.