Structured data extraction uses a language model to pull specific fields out of messy text and return them in a fixed format, such as JSON, that software can use.
In plain words
You first define the shape you want, for example a person with a name and an age, as a schema. Then you send the text and the schema to the model. With a feature such as OpenAI's Structured Outputs, a finished reply is held to that schema unless the model refuses, though the docs say a reply can still fail to match in some cases. Libraries like Instructor wrap this up with Pydantic models in Python and check the result for you.
A quick example
Imagine a pile of customer emails. For each one you ask for the sender's name, the product mentioned and whether they want a refund, as a yes or no field. The model returns one tidy JSON record per email, ready to load into a spreadsheet or database.
Where you'll meet it
- Turning emails, forms and web pages into database rows.
- Feeding clean data to other tools and agents.
Related
Sources
- DocsStructured model outputs, OpenAI API docs · read 29 Sept 2026
- Docsinstructor, 567 Labs (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 11 Nov 2026. Spotted an error? Report it.