How do I make sure the AI always returns valid JSON?

Attach a structured output to your workflow. It's a JSON schema, and every response is checked against it before it reaches your app. If a model returns empty or invalid output, ModelRiver tries your backup models, so your code only ever receives JSON in the shape you defined, or a clear error.

How do I set one up?#

1. Go to Structured Outputs in the console and create a schema, for example summary (string), sentiment (enum) and priority (number).
2. Add sample data. It's what test mode returns.
3. Attach the schema to one or more workflows.

Will ModelRiver fill in missing fields?#

No, never. If a field is missing or a value is the wrong type, that attempt fails and the next backup model is tried. If every attempt fails, you get status: "error". ModelRiver never invents data to make a response pass.

Are there schema features I can't use?#

• pattern and patternProperties aren't accepted.
• format (like email or date) is accepted but not enforced, so validate it in your app if it matters.

What if a model keeps failing the schema?#

• Simplify the schema: fewer nested objects, clearer field names.
• Describe each field in the schema or in the system instructions.
• Use a more capable model as the primary or as a backup.
• Look at the raw response in Request Logs to see what the model actually returned.

Learn more#

• Structured outputs
• Type-safe structured outputs