Do I need structured outputs for extraction use cases?

Not always — but structured outputs are usually the better default for extraction.

When you should use structured outputs

Use them if you need:

• Reliable fields in a fixed schema
  (e.g. name, date, total, address, line items)
• Machine-readable output you can parse without guessing
• Consistency at scale across many documents
• Validation against types, enums, required/optional fields
• Less post-processing and fewer brittle regex rules

This is especially useful for:

• invoices
• receipts
• contracts
• support tickets
• resumes
• medical notes
• web-page/data scraping into a known schema

When you might not need them

You may not need structured outputs if:

• the task is open-ended summarization
• you’re doing light extraction where occasional formatting drift is acceptable
• you only need one or two values and can tolerate some cleanup
• you want the model to produce free-form text first, then manually parse it

Practical rule of thumb

• If the result will go into a database, pipeline, or API, use structured outputs.
• If the result is mainly for a human to read, plain text may be enough.

Example

If you ask:

“Extract the invoice total, due date, and vendor name.”

Plain text might come back as:

“Vendor: Acme Corp. Total is $1,240. Due on March 15.”

Structured output would be:

{
  "vendor_name": "Acme Corp",
  "total": 1240,
  "due_date": "2026-03-15"
}

That’s much easier to use programmatically.

If you want, I can also help you decide when JSON mode is enough vs when to use full schema-enforced structured outputs.