Structured Output
Structured output is a mode where the model is forced to answer in a machine-readable format (JSON, a table, a fixed schema) instead of free text, so its answer can be validated, stored and passed to other software without a human retyping it.
No jargon, no math
In plain words, what is it?
Free text is for humans; software wants fields. Structured output bridges the two. You declare a schema first: field names, types, allowed values ("quantity is a number, grade is one of these five codes"). The schema travels with the prompt as a hard constraint, not a suggestion. As the model generates, the sampler restricts its choices at every step: it can only pick tokens that keep the JSON valid, so a missing bracket or a stray comma is impossible by construction, not by luck. When the generation ends, a parser validates the result against the schema. Valid output goes straight into your ERP, your database, your next tool call. Invalid output triggers a repair pass: the broken piece is sent back to the model with the error message for a fix. The boundary to respect: structure does not create information. If the source document is missing a value, a well-built schema leaves the field empty and flags it; a naive one lets the model fill the gap with a confident guess.
- 01
Define. You declare the output shape: field names, types, allowed values.
- 02
Send. The schema travels with the prompt as a hard constraint.
- 03
Generate. At each step the model can only pick tokens that keep the JSON valid.
- 04
Validate. A parser checks the finished output against the schema.
- 05
Repair or deliver. Invalid output goes back for a fix; valid output flows into your software.
Terms that live nearby.
Function Calling
Function calling is the structured form of tool use: the model emits a machine-readable request naming a function and its arguments, written in JSON, which your code validates and executes, then feeds the return value back into the conversation.
ReadTool Use
Tool use is the capability of an AI model to do more than write text: it can call external tools, such as searching a database, sending an email or running a calculation, by emitting a structured request that software executes for real.
ReadGuardrails
Guardrails are the checks and limits wrapped around an AI model so its behaviour stays inside your rules: they screen what goes in, constrain what the model can do, and verify what comes out before it reaches a person or a system.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.