Mind the layers: a three-layer model for document AI
In case you missed the previous episode: I argued that there is no single universal schema for what a document means , because what matters about a document is relative to the question you are asking. Knowledge is structured by use, the categories we extract are tools rather than universal truths, and the design move that follows is to equip the inquiry rather than model the document. That was…
In this three-part series on document AI, the author introduces a three-layer model for extracting knowledge from documents. The model consists of three layers, each with its own reuse profile and epistemic status.
Layer 1 is the document's intrinsic structure, which includes pages, blocks, tables, reading order, sections, signatures, and geometry. This layer is fully reusable, as every document shares the same basic structure. Its epistemic status is perception, answering the question of what is physically on the page.
Layer 2 consists of domain-specific entities and relations, such as parties, dates, amounts, issuing authorities, and cross-references. This layer is partially reusable, with a sparse upper ontology that generalizes across domains but extends with domain-specific elements. Its epistemic status is grounding, answering the question of what the document is about in domain terms.
Layer 3 is workflow-specific knowledge, such as determining whether a payment is duplicate, assessing the enforceability of a clause, or summarizing a filing for a board. This layer is not reusable and should not be skipped, as it answers the question of what the specific workflow needs to conclude.
The author emphasizes the importance of building the system in layers, with each layer having a unique reuse profile and epistemic status. By separating these layers, the system can be debugged and maintain its accuracy in production. The model applies to various document types, such as legal contracts, invoices, and novels, revealing the same three layers and their distinct roles in knowledge extraction.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
