News for hackers. Stuff that matters.
Re:Launch HN: Reducto Studio (YC W24) – Build accurate docum... by shuss on Tuesday June 24, 2025 @01:10AM (#44363009)
Using LLMs to extract data from documents that don't have a lot of variants is relatively easy with LLMs and (insert your favorite LLM framework here). Things start to get impossible in production when a document you're trying to process has unlimited variants (say, contracts). We recognized the need for this long back and we never released a version of our open source Unstract[1] without a dedicated prompting / JSON schema mapping environment, Prompt Studio[2]. So, welcome to the party.

A lot of folks think they'll write a bunch of prompts in an agents that deal with complex documents. Except, it never works in production with real world data. This is where you need mapping platforms like Unstract (or the competition) that help you modularize this complexity. And the complexity is real. As document variants increase or you need to process documents that need a lot of business logic knowledge to decipher, agents with simple prompts will not work. If you make them work with complex prompts somehow, good luck maintain such agents.

[1] https://github.com/Zipstack/unstract [2] https://docs.unstract.com/unstract/unstract_platform/feature...

There are no comments on this comment yet.