Document AI Glossary

A document AI glossary is a curated vocabulary of the field: the technologies (OCR, layout analysis, vision-language models), the tasks (classification, extraction, summarization, question answering), the operational machinery (confidence thresholds, review queues, straight-through processing), the evaluation language (character error rate, field-level accuracy, faithfulness), and the governance concepts (provenance, residency, audit trails) that together constitute how practitioners talk about turning documents into working data. This site's glossary — the one you are reading — is exactly that, with each term defined in context and linked to its neighbors.

Shared vocabulary earns its keep in this field because the terminology is genuinely treacherous. Near-synonyms carry real distinctions: OCR reads text, parsing recovers structure, extraction produces target fields, understanding spans all of it; IDP names the workflow discipline while document AI names the technology family; agents differ from copilots by autonomy, and both from pipelines by adaptivity. Buying conversations go wrong on these distinctions — a vendor's "99% accuracy" might be character-level OCR accuracy on clean scans while the buyer hears field-level extraction accuracy on their fax traffic. Precise terms are how requirements, benchmarks, and contracts stay honest.

A glossary also maps the field's structure for newcomers: the pipeline view (capture → preprocess → recognize → understand → extract → validate → route → deliver), the model landscape beneath it, and the operational and compliance layers around it. Used that way, it is less a dictionary than an index into the discipline — each entry a doorway into one component of systems that, assembled well, read documents the way institutions need them read: accurately, at scale, and with evidence.