Infer

Inference turns a prompt into a single structured output. It is the narrow capability narrow capability: one schema-checked answer, no durable transcript, no state. When work needs multi-step reasoning across sessions and domain tools, that is an agent, not an inference.

One-shot inference

A single model call that returns one structured output — classify a record, extract fields from a document, or summarize a row.

Tool-using inference

A call may declare authored `tools` closures, and the model then runs a bounded loop — call, observe, continue — until it closes with the schema-constrained answer. Host tools bound to the invocation may be declared too; without that binding they are refused. The result is still one structured output.

Structured output

The result is validated against a Effect Schema schema, so callers get typed data — never freeform text they must parse.

Inference in action

One call in, one structured value out — a schema-checked result, ready to type:

import { Schema } from 'effect';
import { defineQueryHandler } from '@norbital-ai/bolt/authoring';

const ReviewOutcome = Schema.Struct({
  approved: Schema.Boolean,
  reason: Schema.String,
  suggested_priority: Schema.Literals(['low', 'medium', 'high'])
});

export default defineQueryHandler({
  description: 'Reviews one purchase request.',
  schema: Schema.Struct({ request_id: Schema.String }),
  handler: ({ request_id }, api) =>
    Effect.gen(function* () {
      // One call in, one structured value out — validated, typed, never freeform
      return yield* api.infer({
        schema: ReviewOutcome,
        prompt: 'Review this purchase request.'
      });
    })
});

The ai facility port

Model access flows through the ai facility port: Bolt defines the contract, the host binds the concrete provider at runtime, and tenant code never sees credentials. See Facilities for the full port list.

Sessions are data

Each conversation is one tenant-owned aggregate: conversation with its ordered conversation_message rows, turn attempts, and turn_usage records — transcript, status and usage sync atomically like any other collection.

Infer vs Agent

Infer is one prompt in, one structured output out. Norbius is the agentic workload: a loop that reasons across many steps, calls tools, reads and writes collections, and can draft source. Norbius uses inference under the hood — but it is a different surface to author and a different runtime to run.