From Ingestion to Agents: How AI Teams Build on Document Intelligence — Adit Abraham, Reducto
Source
youtube.com
Author
AI Engineer
Date
Context
Document extraction needs to preserve what a page says, including its mistakes. A talk about Reducto's document-processing workflow describes a failure mode in which an image-and-text model rewrites a page and silently recalculates an incorrect total. That produces a different document, even if the arithmetic improves.
The workflow instead applies targeted corrections to recognized text, such as distinguishing a zero from the letter O. The intended benefit is to repair reading errors without inviting a fresh rewrite of the page. For an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → using extracted figures in later steps, this separates faithful transcription from the separate decision to question a document's contents.
Key takeaways · AI-distilled
Adit Abraham of Reducto argues that agents raise the cost of bad parsing: in RAGRetrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.Full definition → a bad parse cost one answer, but in an agent the error compounds across every later step.
Reducto's agentic OCR applies tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →-level corrections, such as a zero for an O, instead of having a VLM rewrite the page, because a VLM asked to rewrite OCR may helpfully 'fix' a total the human got wrong.
Simple tables go to markdown and complex ones to HTML, but Reducto also renders tables in natural language for retrieval, since embeddingA list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.Full definition → models struggle to match a question like 'how did revenue change' to a blob of tags.
Per the talk, feeding parsed structure instead of raw PDFs lifted other frontier models past the newest one on a PDF-decision benchmark and cut reasoning tokens.
He says frontier models on their own tend to be precise but silently drop rows, while document services show the opposite tradeoff; agent harnesses with a code interpreter and repeated self-checks are pitched as getting both.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
videoYour Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
videoBuild Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI
videoDon't Vibe Code Your Agent Infrastructure - Akshat Bubna, Modal
videoHow to Get Frontier Coding Results at 70% Lower Cost - Nouha Dziri, Cohere
