Computational complexity of schema-guided document extraction
13 points by sidmanchkanti21
by graemefawcett
0 subcomment
We solve that problem in a different manner. Our knowledge system allows the user to define the AST token based pattern for the shape of the text to extract from the documents. This is applied on first read and creates a queryable side channel interface to the document that acts both as attention surface for understanding and as an output surface (COW, emits an edge layer over the source document on charge) for agency.
It's like a regex but applied to the structure of the document itself that generates its own control plane.
The LLM has full access to the document or to the control plane, in either mode so we can target specific meanings to focus context or let it explore to find new patterns
by chaps
0 subcomment
This is a topic that interests me deeply! But dang this article needs some supporting documents/visualizations to get its point across better.
Hacker News