A PDF is a printing format — text is glyphs with coordinates, tables are line segments, reading order is never stored. Your agent can't read that raw. Pick your regime and get the stack.
Built on Jerry Liu's talk "Building the Document Context Layer for AI Agents" (AI Engineer Paris, Sept 2026) — co-founder & CEO of LlamaIndex. All regime guidance comes from the talk; the picker logic is our interactive reading of it.
Three inputs. The recommendation updates live — every move shows you what Jerry would point at.
The talk's core design claim — why one parser is never enough.
Cheap specialized parsers vs. frontier VLMs. Heuristic pipelines (PyPDF, PyMuPDF — cluster text fragments, find tables and paragraphs) are fast and cheap. One-shot VLM reads see visual structure well but hallucinate on text pages, cost more, and miss enterprise document semantics.
The hybrid: a pipeline-based deep understanding of the file container, plus vision for whatever needs visual reasoning — an agent routes automatically between the cheap specialized tier and the advanced tier, and escalates only the pages that deserve it.
Tap to see what a document actually looks like to your parser.
Text is glyphs with coordinates. Tables are line segments with characters at cell positions. Reading order is never stored anywhere. That's the job the parsing layer does before your agent ever sees a word.
1. Parsing layer — digitizes PDFs/Office files into token-efficient, agent-actionable context (OCR, markup loading).
2. Semantic storage layer — document management for humans and agents: an agent interface to capture, store, and manage documents.
3. Agent-level workflows — repeatable, tunable pipelines coded in the platform (invoice processing, KYC, claims), tuned for cost/accuracy — not handed to a generic agent. A million invoices become structured DB rows with citations and confidence scores.
2,000 human-verified pages measuring tables, charts, content validity/faithfulness, and semantic formatting — optimized for how AI agents understand documents, not syntactic correctness. ~50 models and OCR systems benchmarked. His caveat: "definitely not 100% resolved" — benchmark complexity can still grow.