Structured output

intelligence.object asks for output matching a schema and validates it before returning:

import { z } from 'zod'

const Sentiment = z.object({
  sentiment: z.enum(['positive', 'negative', 'neutral']),
  confidence: z.number().min(0).max(1),
})

const result = await intelligence.object(Sentiment, 'I absolutely love this library')
// result is typed as { sentiment: 'positive' | ...; confidence: number }

Because the result is validated against your schema, you can pass it straight into the rest of your program. If the model’s output does not match the schema, object throws instead of returning bad data.

Under the hood this forces a synthetic tool call matching your schema, which avoids per-vendor JSON-mode differences. The consequence is that the model must support tool calling. For a model that writes tool calls as text, wrap its provider with textToolCalls.

Routing

router from @intellibrix/structures does exactly this: one structured classification step, then dispatch to the matching brick. Use it to pick a path from a model’s classification.