Structured outputs

Return valid JSON or enforce a JSON Schema with Chat Completions.

Use response_format when your application needs machine-readable output. Prism supports JSON mode and JSON Schema on POST /v1/chat/completions.

JSON mode

JSON mode guarantees valid JSON but does not enforce a specific shape.

const response = await client.chat.completions.create({
  model: "prism-glm53",
  messages: [
    {
      role: "user",
      content:
        "Respond in JSON with the bug category and a one-sentence explanation: ...",
    },
  ],
  response_format: { type: "json_object" },
});

const result = JSON.parse(response.choices[0].message.content ?? "{}");

Tell the model to return JSON in the prompt. Without that instruction, the model may spend the token budget generating whitespace.

JSON Schema

Use strict JSON Schema when downstream code requires known fields:

const response = await client.chat.completions.create({
  model: "prism-glm53",
  messages: [
    {
      role: "user",
      content: "Classify this pull request and summarize the primary risk: ...",
    },
  ],
  response_format: {
    type: "json_schema",
    json_schema: {
      name: "pull_request_review",
      strict: true,
      schema: {
        type: "object",
        properties: {
          risk: {
            type: "string",
            enum: ["low", "medium", "high"],
          },
          summary: { type: "string" },
          requires_human_review: { type: "boolean" },
        },
        required: ["risk", "summary", "requires_human_review"],
        additionalProperties: false,
      },
    },
  },
});

Example content:

{
  "risk": "high",
  "summary": "The migration drops a populated column before backfilling its replacement.",
  "requires_human_review": true
}

Validate in your application

Schema-constrained generation reduces invalid responses, but it does not replace application validation or authorization.

import { z } from "zod";

const PullRequestReview = z.object({
  risk: z.enum(["low", "medium", "high"]),
  summary: z.string(),
  requires_human_review: z.boolean(),
});

const content = response.choices[0].message.content;
if (content == null) {
  throw new Error("Prism returned no structured content");
}

const review = PullRequestReview.parse(JSON.parse(content));

Keep schemas small and explicit. Deep nesting, large enums, and ambiguous field descriptions make the model’s job harder.

Tool schemas and response schemas solve different problems. A tool schema constrains a function call. response_format constrains the assistant’s final response.