Anvia

Structured output ​

Structured output turns a model response into application data that has crossed a local schema-validation boundary.

text
Model response -> JSON parsing -> schema validation -> application data

A provider may use the schema to guide generation, but product code should trust only the value produced after local validation.

1. Define the contract ​

Anvia uses Zod schemas for agent output schemas, extractors, tool inputs and outputs, and pipeline inputs. Parsed completions (generateCompletion() / streamCompletion()) accept any Standard Schema: Zod, Valibot, ArkType, and other implementing libraries (see Schema design):

ts
import { z } from 'zod'

const ticketSchema = z.object({
  category: z.enum(['billing', 'technical', 'account']),
  priority: z.enum(['low', 'normal', 'high']),
  summary: z.string().min(1),
  needsHumanReview: z.boolean(),
})

type Ticket = z.infer<typeof ticketSchema>

Keep the schema small, explicit, and aligned with the decision the application needs to make.

2. Choose where validation happens ​

Use generateCompletion() when one model call should produce one validated value:

ts
import { generateCompletion } from '@anvia/core'

const result = await generateCompletion({
    prompt: message,
    model,
    outputSchema: ticketSchema,
    instructions: 'Classify the support request.'
})

console.log(result.output.priority)

Use an agent outputSchema when tools, retrieval, memory, or multiple turns are needed before the final object. Anvia validates that schema before returning a response outcome; response.output is already typed.

Use extract() when structured fields already exist in source text and a required submission tool is more appropriate than provider-native output schemas.

3. Continue through the section ​

Use a regular completion or prose agent response when only a person will read the answer and no application code needs to branch on its fields.

Built for Anvia.