Skip to Content
📖 Guide DocumentsLLM Function CallingstructuredOutput()

typia.llm.structuredOutput — schema + parse + coerce + validate, in one shot

When you ask an LLM for structured output (one specific JSON shape, no function selection involved), you typically need four things:

  1. The schema to put in response_format / json_schema
  2. A parser that copes with messy LLM output
  3. A coercer for SDKs that already JSON-parsed
  4. A validator to confirm the result actually matches your type

typia.llm.structuredOutput<T>() returns all four bundled together for a single TypeScript type T.

signature
export namespace llm {
  function structuredOutput<
    T extends Record<string, any>,
    Config extends Partial<ILlmSchema.IConfig & { equals: boolean }> = {},
  >(): ILlmStructuredOutput<T>;
}
what-you-get
interface ILlmStructuredOutput<T> {
  parameters: ILlmSchema.IParameters;             // hand this to the LLM
  parse: (input: string) => IJsonParseResult<T>;  // raw string in
  coerce: (input: unknown) => T;                  // pre-parsed object in
  validate: (input: unknown) => IValidation<T>;   // check constraints
}

First example

hello-structured.ts
import typia from "typia";
 
interface IMember {
  email: string;
  name: string;
  age: number;
  hobbies: string[];
}
 
const output = typia.llm.structuredOutput<IMember>();
 
// hand `output.parameters` to your LLM SDK as JSON schema
// then:
const parsed = output.parse(rawLlmText);
if (parsed.success) {
  const v = output.validate(parsed.data);
  if (v.success) doSomething(v.data);
}

undefined

typia
export namespace llm {
  export function structuredOutput<
    T extends Record<string, any>,
    Config extends Partial<ILlmSchema.IConfig & { equals: boolean }> = {},
  >(): ILlmStructuredOutput<T>;
}
examples/src/llm/structuredOutput.ts
import { LlmJson } from "@typia/utils";
import OpenAI from "openai";
import typia, { tags } from "typia";
 
interface IMember {
  email: string & tags.Format<"email">;
  name: string;
  age: number & tags.Minimum<0>;
  hobbies: string[];
  joined_at: string & tags.Format<"date">;
}
 
const main = async (): Promise<void> => {
  // Generate structured output interface
  const output = typia.llm.structuredOutput<IMember>();
 
  // Use schema with OpenAI
  const client: OpenAI = new OpenAI({
    apiKey: "<YOUR_OPENAI_API_KEY>",
  });
  const completion: OpenAI.ChatCompletion =
    await client.chat.completions.create({
      model: "gpt-4o",
      messages: [
        {
          role: "user",
          content: [
            "I am a new member of the community.",
            "",
            "My name is John Doe, and I am 25 years old.",
            "I like playing basketball and reading books,",
            "and joined to this community at 2022-01-01.",
          ].join("\n"),
        },
      ],
      response_format: {
        type: "json_schema",
        json_schema: {
          name: "member",
          schema: output.parameters as any,
        },
      },
    });
 
  // Parse LLM response with type coercion
  const parsed = output.parse(completion.choices[0].message.content!);
  if (!parsed.success) {
    console.error("Parse failed:", parsed.errors);
    return;
  }
 
  // Validate the parsed data
  const validated = output.validate(parsed.data);
  if (!validated.success) {
    // Format errors for LLM feedback
    console.error(LlmJson.stringify(validated));
    return;
  }
 
  console.log("Success:", validated.data);
};
main().catch(console.error);

When to use this

If you need…Use
One JSON shape with parse + coerce + validatestructuredOutput<T>()
One JSON shape, schema only (no helpers)parameters<T>()
Multiple functions the LLM picks betweenapplication<Class>()
Schema of an inner sub-type (advanced)schema<T>()

Parse and coerce

The same parse / coerce / validate / stringify harness as application:

raw JSON string from the LLM
const output = typia.llm.structuredOutput<IMember>();
 
const jsonString = '{"name": "John", "age": "25"}';
 
const result = output.parse(jsonString);
if (result.success) {
  console.log(result.data.age); // 25 — coerced to number
}

Type coercion handles (full list in LlmJson):

  • "42" → 42 for number/integer fields
  • "true" / "false" → true / false for boolean fields
  • "null" → null for nullable fields
  • "{...}" → object (double-stringified objects)
  • "[...]" → array (double-stringified arrays)
  • Nested coercion descends into properties recursively

parse vs coerce — same rule as in application.

  • Raw string from the LLM → parse(text)
  • Already-parsed object from the SDK → coerce(obj)

parse is coerce with a lenient JSON parser in front.

Validation feedback

feedback-loop.ts
import { LlmJson } from "@typia/utils";
 
const output = typia.llm.structuredOutput<IMember>();
const parsed = output.parse(llmResponse);
 
if (parsed.success) {
  const v = output.validate(parsed.data);
 
  if (!v.success) {
    // Annotated JSON the LLM can read and self-correct from
    const feedback = LlmJson.stringify(v);
    // Send `feedback` back to the model in the next turn
  }
}

The formatted output:

{
  "name": "John",
  "age": -5, // ❌ [{"path":"$input.age","expected":"number & Minimum<0>"}]
  "email": "invalid", // ❌ [{"path":"$input.email","expected":"string & Format<\"email\">"}]
}

Same pattern, same payoff as in application — qwen3-coder-next jumps from 6.75% raw success to 100% with the harness around it.

Strict mode

const strict = typia.llm.structuredOutput<IMember, { equals: true }>();
 
strict.validate({
  name: "John",
  age: 25,
  email: "john@example.com",
  hobbies: ["reading"],
  extraField: "not allowed", // ❌ rejected
});

{ equals: true } switches the validator to validateEquals semantics. (The schema-side additionalProperties: false is always emitted on LLM parameter schemas, regardless of the flag — equals only changes how the runtime validator treats stray properties at validation time.)

Restrictions

Same as parameters and schema:

  • T must be a keyworded object type with static keys
  • No dynamic keys (Record<string, V>)
  • T may not be nullable or optional
src/examples/llm.structuredOutput.violation.ts
import typia from "typia";
 
typia.llm.structuredOutput<string>();
typia.llm.structuredOutput<Record<string, boolean>>();
typia.llm.structuredOutput<Array<number>>();

Where to go next

  • Multiple functions, LLM picks → application
  • Just the schema, no helpers → parameters
  • Harness internals (parse, coerce, validate, stringify) → LlmJson
Last updated on
application() functionparameters() function