Structured Document Extraction with LLMs - From First Principles to Production

Structured generation with LLMs makes it easier to extract standardized data from complex documents. But the process for building & evaluating the right schema is still largely artisanal. Retab has built an end-to-end platform for rigorously building these schema-driven pipelines. In this talk, Louis will go over technical insights and best practices.

He will cover the whole process:

• Going from business problem to JSON schema
• Context engineering for large documents
• Designing evals for first-order and second-order extractions
• Using k-LLM consensus to boost performance and quantify uncertainty
• Retab’s agent for auto-optimizing JSON schema descriptions
• Human-in-the-loop for schema-driven automations