Usage

Basic Usage

Instructor extracts structured data from text using LLM inference. You define a PHP class that describes the shape of the data you want, and Instructor takes care of building the prompt, calling the model, and deserializing the response into a typed object.

use Cognesy\Instructor\StructuredOutput;

class Person {
    public string $name;
    public int $age;
}

$person = (new StructuredOutput)
    ->with(
        messages: 'Jason is 28 years old.',
        responseModel: Person::class,
    )
    ->get();

echo $person->name; // Jason
echo $person->age;  // 28
// @doctest id="00b1"

By default, Instructor looks for the OPENAI_API_KEY environment variable. You can also choose a provider explicitly with StructuredOutput::using('openai') or by passing a runtime configured with LLMConfig.

Building The Request

The with() method covers the common path. It accepts all the parameters you typically need in a single call:

$person = (new StructuredOutput)
    ->with(
        messages: 'Jason is 28 years old.',
        responseModel: Person::class,
        system: 'Extract accurate data.',
        prompt: 'Identify the person mentioned.',
        model: 'gpt-4o',
    )
    ->get();
// @doctest id="15b9"

When you prefer a more explicit, step-by-step style, use the fluent API:

$person = (new StructuredOutput)
    ->withMessages('Jason is 28 years old.')
    ->withResponseModel(Person::class)
    ->withSystem('Extract accurate data.')
    ->withPrompt('Identify the person mentioned.')
    ->withModel('gpt-4o')
    ->get();
// @doctest id="89e5"

Both approaches produce identical requests. Use whichever reads better in your code.

Request Methods

Method	Purpose
`withMessages(...)`	Set the chat messages
`withInput(...)`	Set input from a string, array, or object (converted to messages)
`withResponseModel(...)`	Set the response model (class string, instance, or schema array)
`withResponseClass(...)`	Set the response model from a class name
`withResponseObject(...)`	Set the response model from an object instance
`withResponseJsonSchema(...)`	Set the response model from a JSON Schema array
`withSystem(...)`	Set the system prompt (`string\|\Stringable`)
`withPrompt(...)`	Set additional prompt text (`string\|\Stringable`)
`withExamples(...)`	Provide few-shot examples
`withModel(...)`	Override the model name
`withOptions(...)`	Pass provider-specific options
`withOption(...)`	Set a single provider option
`withStreaming(...)`	Enable or disable streaming
`withCachedContext(...)`	Set cached context for providers that support prompt caching

Reading The Result

Instructor provides several ways to consume the response depending on your needs.

`get()` - The Parsed Value

The most common method. Returns the deserialized, validated object (or scalar when using the Scalar adapter):

$person = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->get();
// @doctest id="2f30"

`response()` - The Full Response Envelope

Returns a StructuredOutputResponse that wraps both the parsed value and the raw LLM response, giving you access to usage metadata, finish reason, and more:

$response = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->response();

$person = $response->value();
$usage  = $response->usage();
// @doctest id="f323"

`inferenceResponse()` - The Underlying Inference Response

Returns the low-level InferenceResponse from the Polyglot layer, useful when you need direct access to HTTP response data or provider-specific details:

$raw = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->inferenceResponse();
// @doctest id="b676"

`stream()` - Streaming Partial Results

Returns a StructuredOutputStream for real-time processing. Streaming is enabled automatically when you call stream():

$stream = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->stream();

foreach ($stream->partials() as $partial) {
    echo $partial->name ?? '...';
}

$person = $stream->lastUpdate();
// @doctest id="5268"

`create()` - Lazy Execution

Returns a PendingStructuredOutput handle without triggering the LLM call. Nothing executes until you read from it:

$pending = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->create();

// execution happens here
$person = $pending->get();
// @doctest id="6b80"

PendingStructuredOutput exposes the same reading methods as StructuredOutput plus a few utility helpers:

Method	Return type
`get()`	The parsed value
`response()`	`StructuredOutputResponse`
`inferenceResponse()`	`InferenceResponse`
`stream()`	`StructuredOutputStream`
`toJson()`	JSON string of the extracted data
`toArray()`	Associative array of the extracted data
`toJsonObject()`	`Json` object

Typed Convenience Methods

When working with Scalar responses or any result where you know the expected PHP type, you can skip get() and call a typed accessor directly:

$age = (new StructuredOutput)
    ->with(messages: 'Jason is 28.', responseModel: Scalar::integer('age'))
    ->getInt();
// @doctest id="35eb"

Available typed methods: getString(), getInt(), getFloat(), getBoolean(), getObject(), getArray().

String As Input

You can pass a plain string anywhere messages are expected. Instructor wraps it into a user message automatically:

$person = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->get();
// @doctest id="b676"

This is equivalent to passing [['role' => 'user', 'content' => 'Jason is 28 years old.']].

Structured-To-Structured Processing

The input parameter accepts objects, arrays, or strings. This lets you transform one structured representation into another:

class Email {
    public function __construct(
        public string $address = '',
        public string $subject = '',
        public string $body = '',
    ) {}
}

$email = new Email(
    address: 'joe@gmail.com',
    subject: 'Status update',
    body: 'Your account has been updated.',
);

$translated = (new StructuredOutput)
    ->withInput($email)
    ->with(
        responseModel: Email::class,
        prompt: 'Translate the text fields to Spanish. Keep other fields unchanged.',
    )
    ->get();
// @doctest id="e2b6"

Output Formats

By default, Instructor returns an instance of your response model class. You can change this with the output format methods:

// Return as an associative array instead of an object
$data = (new StructuredOutput)
    ->withResponseClass(User::class)
    ->intoArray()
    ->with(messages: 'John Doe, 30 years old')
    ->get();
// ['name' => 'John Doe', 'age' => 30]

// Use one class for the schema but hydrate into a different class
$dto = (new StructuredOutput)
    ->withResponseClass(UserProfile::class)
    ->intoInstanceOf(UserDTO::class)
    ->with(messages: 'Extract user data')
    ->get();
// @doctest id="7ff8"

Three output format methods are available:

Method	Effect
`intoArray()`	Skip deserialization, return a raw associative array
`intoInstanceOf($class)`	Use the schema from the response model but hydrate into a different class
`intoObject($obj)`	Pass a self-deserializing object that implements `CanDeserializeSelf`

Using A Runtime

For applications that share provider configuration and behavior across many requests, create a StructuredOutputRuntime once and reuse it:

use Cognesy\Instructor\StructuredOutput;
use Cognesy\Instructor\StructuredOutputRuntime;
use Cognesy\Polyglot\Inference\Config\LLMConfig;

$runtime = StructuredOutputRuntime::fromConfig(
    LLMConfig::fromPreset('openai')
)->withMaxRetries(2);

$person = (new StructuredOutput)
    ->withRuntime($runtime)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->get();
// @doctest id="a9e4"

The runtime holds settings like retries, output mode, validators, transformers, and deserializers. Individual requests stay lightweight and focused on content. You can also use the static shorthand to pick a provider without building a full runtime:

$person = StructuredOutput::using('anthropic')
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->get();
// @doctest id="b0ff"

Streaming Support

Instructor supports streaming of partial results, allowing you to process data as it arrives from the model:

$stream = (new StructuredOutput)
    ->with(messages: 'Jason is 28 years old.', responseModel: Person::class)
    ->stream();

foreach ($stream->partials() as $partialPerson) {
    echo "Name: " . ($partialPerson->name ?? '...');
    echo "Age: " . ($partialPerson->age ?? '...');
}

// After the stream completes, retrieve the final validated object
$person = $stream->lastUpdate();
// @doctest id="0b9b"

The StructuredOutputStream provides several iteration methods:

Method	Yields
`partials()`	Partially filled objects as they arrive
`sequence()`	Completed items when using `Sequence` as the response model
`responses()`	Full `StructuredOutputResponse` snapshots
`finalValue()`	Drains the stream and returns the final parsed value
`finalResponse()`	Drains the stream and returns the final `StructuredOutputResponse`

Packages

Instructor

Polyglot

Agents

Agent-ctrl

Telemetry

Sandbox

HTTP Client

Laravel

Symfony

Xprompt

Basic Usage

Building The Request

Request Methods

Reading The Result

`get()` - The Parsed Value

`response()` - The Full Response Envelope

`inferenceResponse()` - The Underlying Inference Response

`stream()` - Streaming Partial Results

`create()` - Lazy Execution

Typed Convenience Methods

String As Input

Structured-To-Structured Processing

Output Formats

Using A Runtime

Streaming Support

​Basic Usage

​Building The Request

​Request Methods

​Reading The Result

​get() - The Parsed Value

​response() - The Full Response Envelope

​inferenceResponse() - The Underlying Inference Response

​stream() - Streaming Partial Results

​create() - Lazy Execution

​Typed Convenience Methods

​String As Input

​Structured-To-Structured Processing

​Output Formats

​Using A Runtime

​Streaming Support

Basic Usage

Building The Request

Request Methods

Reading The Result

`get()` - The Parsed Value

`response()` - The Full Response Envelope

`inferenceResponse()` - The Underlying Inference Response

`stream()` - Streaming Partial Results

`create()` - Lazy Execution

Typed Convenience Methods

String As Input

Structured-To-Structured Processing

Output Formats

Using A Runtime

Streaming Support