Definition
What is an AI-native document representation (AIDR)?
An AI-native document representation (AIDR) is a deterministic, evidence-anchored, model-independent representation of a document: one upload is transformed once into eight synchronized layers — from canonical text to entity and reasoning graphs — and every AI artifact (markdown, structured JSON, embeddings, search, cited answers) is derived from that representation rather than re-computed by a model on every request.
The term names a shift in where the source of truth lives. In a parser-centric stack, truth lives in whatever the model produced last run — change the model, re-run the corpus, hope nothing drifted. In an AIDR stack, truth lives in the representation: models are interchangeable readers and writers of it, and every derived claim carries its evidence (page, span, bounding box) with it.
Docsfra introduced the term and implements it as the core of its platform; this page defines the concept precisely so it can be evaluated on its own merits.
The eight layers
01
Canonical Document
The normalized, layout-aware reading of the source — one authoritative text with structure.
02
Document Memory
Persistent, versioned state of everything derived from the document over time.
03
Knowledge Units
Typed, self-contained units of meaning extracted from the canonical text.
04
Structured Objects
Schema-shaped data — fields, tables, sets — extracted with their source anchors.
05
Entity Graph
The people, organizations, amounts and dates in the document, linked and reconciled.
06
Reasoning Graph
Claims and the inference chains that support them, each step anchored to evidence.
07
Evidence
Page, span and bounding-box anchors that make every derived statement verifiable.
08
Provenance
Which model, prompt and version produced each derivation — the audit trail.
Why it matters
- Model independence: swapping LLMs or embedding models does not force re-processing — the representation persists, artifacts re-derive.
- Verifiability: answers cite page, span and box; a hallucinated claim has nowhere to hide because every statement must anchor to evidence.
- Determinism: the same document yields the same representation, which makes document AI auditable — a requirement in trade finance, legal and government use.
- Economics: expensive understanding runs once per document, not once per question.
Frequently asked questions
How is an AIDR different from a vector index?
A vector index stores one projection of a document (embeddings) for one purpose (similarity search) tied to one model. An AIDR stores the document's full understood form — text, structure, entities, reasoning, evidence — from which a vector index is just one derivable artifact.
Is AIDR a standard?
It is a concept coined and implemented by Docsfra. The definition on this page is open: any system that produces a deterministic, evidence-anchored, model-independent multi-layer representation qualifies.
Where can I see one?
Docsfra's console shows every layer of a processed document, and the API exposes them at api.docsfra.com — sign up with 250 free credits to inspect a real AIDR built from your own file.