I implemented a similar pattern in my LLM tool and Python library back in February: https://simonwillison.net/2025/Feb/28/llm-schemas/

My version works with Pydantic models or JSON schema in Python code, or with JSON schema or a weird DSL I invented on the command-line:

  curl https://news.ycombinator.com/ | \
    llm --schema-multi 'headline,url,votes int' \
    -m gpt-4.1 --system 'all links'

There is also Unstract(open-source) that helps process structured data extraction. Key differences:

1. Unstract has a Pre-processing layer(OCR). Which converts documents into LLM readable formats.(helps improve accuracy, and control costs)

2. Unstract also connects to your existing data sources, making it an out-of-the-box ETL tool.

https://github.com/Zipstack/unstract

It’s not extracting data _from_ the model it is using the model to extract structured data from the input.

The use case that immediately comes to mind is analysis of legal documents. Lawyers spend a lot of time going through piles of contracts during due diligence for any kind of investment or acquisition transaction, painstakingly identifying concepts that need to be addressed in various ways. LLMs are decent at doing this kind of work, but error-prone (as are humans, by the way). Having a way to visualize the results could be helpful in speeding up the review process of the LLM’s work.

For complex business documents -- one approach was to use Named Entity Recognition to identify all entities and use that to build a knowledge graph to serve as a complementary repository of knowledge (in addition to the vector embeddings of semantic chunks) to aid RAG workflows.

Does this proposed approach complement this or supercede the need for NER / Knowledge Graph. Just wondering aloud. Appreciate any insights here.

Oly Chit! This is a BIG deal! Sub-page citations…in-context RAG…built-in HTML UI…this is like the holy grail of deterministic text extraction. I’m trying this ASAP Rocky.

In the example, if `extraction_class` can be any string, how does it know that "relationship" implies it should have attributes "character_1" and "character_2" when your example data didn't?

You could use this to generate character graphs from big novels. Make an app that allows you to input a page number so the model only extracts characters you've encountered thus far.

Does this work with other open-source LLMs like Qwen3 or other OpenAI compatible LLM Apis?

wiring this to wikidata would be great