Skip to content
Getting started with local LLMs (SWIB26)
Part 2: Working with your data
Structured output

Structured output

A catalog needs separate fields: a title, a creator, a place, a year. A chat answer gives you a paragraph. A person can read a paragraph, but a spreadsheet, a record store or a validation script cannot. This chapter makes the model fill a fixed set of fields instead, and then checks every answer. That is what structured output means: an answer in a fixed format that a program can read.

Notebook 02, 02_structured_output, has a 25-minute budget. In the session it gets about 22 minutes, from minute 28 to minute 50 of Part 2. We give the model a catalog card and get a record back. Then we watch it invent a value the card never stated, and we add a rule that catches the invention. Finally, we measure twenty cards against a gold file, which holds the correct answers.

What the notebook covers

  1. Define the shape of the answer as a Pydantic model, a Python class that lists the fields and their types. Hand its JSON schema to the model.
  2. Validate what comes back, and learn what grammar-enforced output does and does not guarantee.
  3. See the model invent a missing field, and add a rule that catches it.
  4. Run twenty cards in a batch and measure field accuracy against a gold file.
A small model will confidently fill a field the card is silent about. Valid JSON is not the same as a true record, so every extraction in this chapter is checked.

The pages

Concepts
JSON schema, Pydantic, enforced decoding and the retry.
Exercise
One card, one invented place, twenty cards measured.

Setup is on the environment page and the notebooks are listed on the notebooks page.