README

Structured data extraction in PHP, powered by LLMs. Designed for simplicity, transparency, and control.

What is Instructor?

Instructor is a library that allows you to extract structured, validated data from multiple types of inputs: text, images or OpenAI style chat sequence arrays. It is powered by Large Language Models (LLMs).

Instructor simplifies LLM integration in PHP projects. It handles the complexity of extracting structured data from LLM outputs, so you can focus on building your application logic and iterate faster.

Instructor for PHP is inspired by the Instructor library for Python created by Jason Liu.

Here's a simple CLI demo app using Instructor to extract structured data from text:

Feature highlights

Core features

Get structured responses from LLMs without writing boilerplate code
Validation of returned data
Automated retries in case of errors when LLM responds with invalid data
Integrate LLM support into your existing PHP code with minimal friction - no framework, no extensive code changes

Flexible inputs

Process various types of input data: text, series of chat messages or images using the same, simple API
'Structured-to-structured' processing - provide object or array as an input and get object with the results of inference back
Demonstrate examples to improve the quality of inference

Customization

Define response data model the way you want: type-hinted classes, JSON Schema arrays, or dynamic data shapes with Structure class
Customize prompts and retry prompts
Use attributes or PHP DocBlocks to provide additional instructions for LLM
Customize response model processing by providing your own implementation of schema, deserialization, validation and transformation interfaces

Sync and streaming support

Supports both synchronous or streaming responses
Get partial updates & stream completed sequence items

Observability

Get detailed insight into internal processing via events
Debug mode to see the details of LLM API requests and responses

Support for multiple LLMs / API providers

Easily switch between LLM providers
Support for most popular LLM APIs (incl. OpenAI, Gemini, Anthropic, Cohere, Azure, Groq, Mistral, Fireworks AI, Together AI)
OpenRouter support - access to 100+ language models
Use local models with Ollama

Other capabilities

Developer friendly LLM context caching for reduced costs and faster inference (for Anthropic models)
Developer friendly data extraction from images (for OpenAI, Anthropic and Gemini models)

Documentation and examples

Learn more from growing documentation and 50+ cookbooks

Instructor in Other Languages

Check out implementations in other languages below:

Python (original)
Javascript (port)
Elixir (port)

If you want to port Instructor to another language, please reach out to us on Twitter we'd love to help you get started!

How Instructor Enhances Your Workflow

Instructor introduces three key enhancements compared to direct API usage.

Response Model

You just specify a PHP class to extract data into via the 'magic' of LLM chat completion. And that's it.

Instructor reduces brittleness of the code extracting the information from textual data by leveraging structured LLM responses.

Instructor helps you write simpler, easier to understand code - you no longer have to define lengthy function call definitions or write code for assigning returned JSON into target data objects.

Validation

Response model generated by LLM can be automatically validated, following set of rules. Currently, Instructor supports only Symfony validation.

You can also provide a context object to use enhanced validator capabilities.

Max Retries

You can set the number of retry attempts for requests.

Instructor will repeat requests in case of validation or deserialization error up to the specified number of times, trying to get a valid response from LLM.

Get Started

Installing Instructor is simple. Run following command in your terminal, and you're on your way to a smoother data handling experience!

composer require cognesy/instructor-php

Usage

Basic example

This is a simple example demonstrating how Instructor retrieves structured information from provided text (or chat message sequence).

Response model class is a plain PHP class with typehints specifying the types of fields of the object.

use Cognesy\Instructor\Instructor;

// Step 0: Create .env file in your project root:
// OPENAI_API_KEY=your_api_key

// Step 1: Define target data structure(s)
class Person {
    public string $name;
    public int $age;
}

// Step 2: Provide content to process
$text = "His name is Jason and he is 28 years old.";

// Step 3: Use Instructor to run LLM inference
$person = (new Instructor)->respond(
    messages: $text,
    responseModel: Person::class,
);

// Step 4: Work with structured response data
assert($person instanceof Person); // true
assert($person->name === 'Jason'); // true
assert($person->age === 28); // true

echo $person->name; // Jason
echo $person->age; // 28

var_dump($person);
// Person {
//     name: "Jason",
//     age: 28
// }

NOTE: Instructor supports classes / objects as response models. In case you want to extract simple types or enums, you need to wrap them in Scalar adapter - see section below: Extracting Scalar Values.

Connecting to various LLM API providers

Instructor allows you to define multiple API connections in llm.php file. This is useful when you want to use different LLMs or API providers in your application.

Default configuration is located in /config/llm.php in the root directory of Instructor codebase. It contains a set of predefined connections to all LLM APIs supported out-of-the-box by Instructor.

Config file defines connections to LLM APIs and their parameters. It also specifies the default connection to be used when calling Instructor without specifying the client connection.

    // This is fragment of /config/llm.php file
    'defaultConnection' => 'openai',
    //...
    'connections' => [
        'anthropic' => [ ... ],
        'azure' => [ ... ],
        'cohere1' => [ ... ],
        'cohere2' => [ ... ],
        'fireworks' => [ ... ],
        'gemini' => [ ... ],
        'grok' => [ ... ],
        'groq' => [ ... ],
        'mistral' => [ ... ],
        'ollama' => [
            'providerType' => LLMProviderType::Ollama->value,
            'apiUrl' => 'http://localhost:11434/v1',
            'apiKey' => Env::get('OLLAMA_API_KEY', ''),
            'endpoint' => '/chat/completions',
            'defaultModel' => 'qwen2.5:0.5b',
            'defaultMaxTokens' => 1024,
            'httpClient' => 'guzzle-ollama', // use custom HTTP client configuration
        ],
        'openai' => [ ... ],
        'openrouter' => [ ... ],
        'together' => [ ... ],
    // ...

To customize the available connections you can either modify existing entries or add your own.

Connecting to LLM API via predefined connection is as simple as calling withClient method with the connection name.

<?php
// ...
$user = (new Instructor)
    ->withConnection('ollama')
    ->respond(
        messages: "His name is Jason and he is 28 years old.",
        responseModel: Person::class,
    );
// ...

You can change the location of the configuration files for Instructor to use via INSTRUCTOR_CONFIG_PATH environment variable. You can use copies of the default configuration files as a starting point.

Structured-to-structured processing

Instructor offers a way to use structured data as an input. This is useful when you want to use object data as input and get another object with a result of LLM inference.

The input field of Instructor's respond() and request() methods can be an object, but also an array or just a string.

<?php
use Cognesy\Instructor\Instructor;

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

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

$translation = (new Instructor)->respond(
    input: $email,
    responseModel: Email::class,
    prompt: 'Translate the text fields of email to Spanish. Keep other fields unchanged.',
);

assert($translation instanceof Email); // true
dump($translation);
// Email {
//     address: "joe@gmail",
//     subject: "Actualización de estado",
//     body: "Su cuenta ha sido actualizada."
// }
?>

Validation

Instructor validates results of LLM response against validation rules specified in your data model.

For further details on available validation rules, check Symfony Validation constraints.

use Symfony\Component\Validator\Constraints as Assert;

class Person {
    public string $name;
    #[Assert\PositiveOrZero]
    public int $age;
}

$text = "His name is Jason, he is -28 years old.";
$person = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => $text]],
    responseModel: Person::class,
);

// if the resulting object does not validate, Instructor throws an exception

Max Retries

In case maxRetries parameter is provided and LLM response does not meet validation criteria, Instructor will make subsequent inference attempts until results meet the requirements or maxRetries is reached.

Instructor uses validation errors to inform LLM on the problems identified in the response, so that LLM can try self-correcting in the next attempt.

use Symfony\Component\Validator\Constraints as Assert;

class Person {
    #[Assert\Length(min: 3)]
    public string $name;
    #[Assert\PositiveOrZero]
    public int $age;
}

$text = "His name is JX, aka Jason, he is -28 years old.";
$person = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => $text]],
    responseModel: Person::class,
    maxRetries: 3,
);

// if all LLM's attempts to self-correct the results fail, Instructor throws an exception

Alternative ways to call Instructor

You can call request() method to set the parameters of the request and then call get() to get the response.

use Cognesy\Instructor\Instructor;

$instructor = (new Instructor)->request(
    messages: "His name is Jason, he is 28 years old.",
    responseModel: Person::class,
);
$person = $instructor->get();

Streaming support

Instructor supports streaming of partial results, allowing you to start processing the data as soon as it is available.

<?php
use Cognesy\Instructor\Instructor;

$stream = (new Instructor)->request(
    messages: "His name is Jason, he is 28 years old.",
    responseModel: Person::class,
    options: ['stream' => true]
)->stream();

foreach ($stream as $partialPerson) {
    // process partial person data
    echo $partialPerson->name;
    echo $partialPerson->age;
}

// after streaming is done you can get the final, fully processed person object...
$person = $stream->getLastUpdate()
// ...to, for example, save it to the database
$db->save($person);
?>

Partial results

You can define onPartialUpdate() callback to receive partial results that can be used to start updating UI before LLM completes the inference.

NOTE: Partial updates are not validated. The response is only validated after it is fully received.

use Cognesy\Instructor\Instructor;

function updateUI($person) {
    // Here you get partially completed Person object update UI with the partial result
}

$person = (new Instructor)->request(
    messages: "His name is Jason, he is 28 years old.",
    responseModel: Person::class,
    options: ['stream' => true]
)->onPartialUpdate(
    fn($partial) => updateUI($partial)
)->get();

// Here you get completed and validated Person object
$this->db->save($person); // ...for example: save to DB

Shortcuts

String as Input

You can provide a string instead of an array of messages. This is useful when you want to extract data from a single block of text and want to keep your code simple.

// Usually, you work with sequences of messages:

$value = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => "His name is Jason, he is 28 years old."]],
    responseModel: Person::class,
);

// ...but if you want to keep it simple, you can just pass a string:

$value = (new Instructor)->respond(
    messages: "His name is Jason, he is 28 years old.",
    responseModel: Person::class,
);

Extracting Scalar Values

Sometimes we just want to get quick results without defining a class for the response model, especially if we're trying to get a straight, simple answer in a form of string, integer, boolean or float. Instructor provides a simplified API for such cases.

use Cognesy\Instructor\Extras\Scalar\Scalar;
use Cognesy\Instructor\Instructor;

$value = (new Instructor)->respond(
    messages: "His name is Jason, he is 28 years old.",
    responseModel: Scalar::integer('age'),
);

var_dump($value);
// int(28)

In this example, we're extracting a single integer value from the text. You can also use Scalar::string(), Scalar::boolean() and Scalar::float() to extract other types of values.

Extracting Enum Values

Additionally, you can use Scalar adapter to extract one of the provided options by using Scalar::enum().

use Cognesy\Instructor\Extras\Scalar\Scalar;
use Cognesy\Instructor\Instructor;

enum ActivityType : string {
    case Work = 'work';
    case Entertainment = 'entertainment';
    case Sport = 'sport';
    case Other = 'other';
}

$value = (new Instructor)->respond(
    messages: "His name is Jason, he currently plays Doom Eternal.",
    responseModel: Scalar::enum(ActivityType::class, 'activityType'),
);

var_dump($value);
// enum(ActivityType:Entertainment)

Extracting Sequences of Objects

Sequence is a wrapper class that can be used to represent a list of objects to be extracted by Instructor from provided context.

It is usually more convenient not create a dedicated class with a single array property just to handle a list of objects of a given class.

Additional, unique feature of sequences is that they can be streamed per each completed item in a sequence, rather than on any property update.

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

$text = <<<TEXT
    Jason is 25 years old. Jane is 18 yo. John is 30 years old
    and Anna is 2 years younger than him.
TEXT;

$list = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => $text]],
    responseModel: Sequence::of(Person::class),
    options: ['stream' => true]
);

See more about sequences in the Sequences section.

Specifying Data Model

Type Hints

Use PHP type hints to specify the type of extracted data.

Use nullable types to indicate that given field is optional.

    class Person {
        public string $name;
        public ?int $age;
        public Address $address;
    }

DocBlock type hints

You can also use PHP DocBlock style comments to specify the type of extracted data. This is useful when you want to specify property types for LLM, but can't or don't want to enforce type at the code level.

class Person {
    /** @var string */
    public $name;
    /** @var int */
    public $age;
    /** @var Address $address person's address */
    public $address;
}

See PHPDoc documentation for more details on DocBlock website.

Typed Collections / Arrays

PHP currently does not support generics or typehints to specify array element types.

Use PHP DocBlock style comments to specify the type of array elements.

class Person {
    // ...
}

class Event {
    // ...
    /** @var Person[] list of extracted event participants */
    public array $participants;
    // ...
}

Complex data extraction

Instructor can retrieve complex data structures from text. Your response model can contain nested objects, arrays, and enums.

use Cognesy\Instructor\Instructor;

// define a data structures to extract data into
class Person {
    public string $name;
    public int $age;
    public string $profession;
    /** @var Skill[] */
    public array $skills;
}

class Skill {
    public string $name;
    public SkillType $type;
}

enum SkillType {
    case Technical = 'technical';
    case Other = 'other';
}

$text = "Alex is 25 years old software engineer, who knows PHP, Python and can play the guitar.";

$person = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => $text]],
    responseModel: Person::class,
); // client is passed explicitly, can specify e.g. different base URL

// data is extracted into an object of given class
assert($person instanceof Person); // true

// you can access object's extracted property values
echo $person->name; // Alex
echo $person->age; // 25
echo $person->profession; // software engineer
echo $person->skills[0]->name; // PHP
echo $person->skills[0]->type; // SkillType::Technical
// ...

var_dump($person);
// Person {
//     name: "Alex",
//     age: 25,
//     profession: "software engineer",
//     skills: [
//         Skill {
//              name: "PHP",
//              type: SkillType::Technical,
//         },
//         Skill {
//              name: "Python",
//              type: SkillType::Technical,
//         },
//         Skill {
//              name: "guitar",
//              type: SkillType::Other
//         },
//     ]
// }

Dynamic data schemas

If you want to define the shape of data during runtime, you can use Structure class.

Structures allow you to define and modify arbitrary shape of data to be extracted by LLM. Classes may not be the best fit for this purpose, as declaring or changing them during execution is not possible.

With structures, you can define custom data shapes dynamically, for example based on the user input or context of the processing, to specify the information you need LLM to infer from the provided text or chat messages.

Example below demonstrates how to define a structure and use it as a response model:

<?php
use Cognesy\Instructor\Extras\Structure\Field;
use Cognesy\Instructor\Extras\Structure\Structure;

enum Role : string {
    case Manager = 'manager';
    case Line = 'line';
}

$structure = Structure::define('person', [
    Field::string('name'),
    Field::int('age'),
    Field::enum('role', Role::class),
]);

$person = (new Instructor)->respond(
    messages: 'Jason is 25 years old and is a manager.',
    responseModel: $structure,
);

// you can access structure data via field API...
assert($person->field('name') === 'Jason');
// ...or as structure object properties
assert($person->age === 25);
?>

For more information see Structures section.

Changing LLM model and options

You can specify model and other options that will be passed to OpenAI / LLM endpoint.

use Cognesy\Instructor\Features\LLM\Data\LLMConfig;
use Cognesy\Instructor\Features\LLM\Drivers\OpenAIDriver;
use Cognesy\Instructor\Instructor;

// OpenAI auth params
$yourApiKey = Env::get('OPENAI_API_KEY'); // use your own API key

// Create instance of OpenAI driver initialized with custom parameters
$driver = new OpenAIDriver(new LLMConfig(
    apiUrl: 'https://api.openai.com/v1', // you can change base URI
    apiKey: $yourApiKey,
    endpoint: '/chat/completions',
    metadata: ['organization' => ''],
    model: 'gpt-4o-mini',
    maxTokens: 128,
));

/// Get Instructor with the default client component overridden with your own
$instructor = (new Instructor)->withDriver($driver);

$user = $instructor->respond(
    messages: "Jason (@jxnlco) is 25 years old and is the admin of this project. He likes playing football and reading books.",
    responseModel: User::class,
    model: 'gpt-3.5-turbo',
    options: ['stream' => true ]
);

Support for language models and API providers

Instructor offers out of the box support for following API providers:

Anthropic
Azure OpenAI
Cohere
Fireworks AI
Groq
Mistral
Ollama (on localhost)
OpenAI
OpenRouter
Together AI

For usage examples, check Hub section or examples directory in the code repository.

Using DocBlocks as Additional Instructions for LLM

You can use PHP DocBlocks (/** */) to provide additional instructions for LLM at class or field level, for example to clarify what you expect or how LLM should process your data.

Instructor extracts PHP DocBlocks comments from class and property defined and includes them in specification of response model sent to LLM.

Using PHP DocBlocks instructions is not required, but sometimes you may want to clarify your intentions to improve LLM's inference results.

/**
 * Represents a skill of a person and context in which it was mentioned. 
 */
class Skill {
    public string $name;
    /** @var SkillType $type type of the skill, derived from the description and context */
    public SkillType $type;
    /** Directly quoted, full sentence mentioning person's skill */
    public string $context;
}

Customizing Validation

ValidationMixin

You can use ValidationMixin trait to add ability of easy, custom data object validation.

use Cognesy\Instructor\Features\Validation\Traits\ValidationMixin;

class User {
    use ValidationMixin;

    public int $age;
    public int $name;

    public function validate() : array {
        if ($this->age < 18) {
            return ["User has to be adult to sign the contract."];
        }
        return [];
    }
}

Validation Callback

Instructor uses Symfony validation component to validate extracted data. You can use #[Assert/Callback] annotation to build fully customized validation logic.

use Cognesy\Instructor\Instructor;
use Symfony\Component\Validator\Constraints as Assert;
use Symfony\Component\Validator\Context\ExecutionContextInterface;

class UserDetails
{
    public string $name;
    public int $age;
    
    #[Assert\Callback]
    public function validateName(ExecutionContextInterface $context, mixed $payload) {
        if ($this->name !== strtoupper($this->name)) {
            $context->buildViolation("Name must be in uppercase.")
                ->atPath('name')
                ->setInvalidValue($this->name)
                ->addViolation();
        }
    }
}

$user = (new Instructor)->respond(
    messages: [['role' => 'user', 'content' => 'jason is 25 years old']],
    responseModel: UserDetails::class,
    maxRetries: 2
);

assert($user->name === "JASON");

See Symfony docs for more details on how to use Callback constraint.

Internals

Lifecycle

As Instructor for PHP processes your request, it goes through several stages:

Initialize and self-configure (with possible overrides defined by developer).
Analyze classes and properties of the response data model specified by developer.
Encode data model into a schema that can be provided to LLM.
Execute request to LLM using specified messages (content) and response model metadata.
Receive a response from LLM or multiple partial responses (if streaming enabled).
Deserialize response received from LLM into originally requested classes and their properties.
In case response contained incomplete or corrupted data - if errors are encountered, create feedback message for LLM and requests regeneration of the response.
Execute validations defined by developer for the data model - if any of them fail, create feedback message for LLM and requests regeneration of the response.
Repeat the steps 4-8, unless specified limit of retries has been reached or response passes validation

Receiving notification on internal events

Instructor allows you to receive detailed information at every stage of request and response processing via events.

(new Instructor)->onEvent(string $class, callable $callback) method - receive callback when specified type of event is dispatched
(new Instructor)->wiretap(callable $callback) method - receive any event dispatched by Instructor, may be useful for debugging or performance analysis

Receiving events can help you to monitor the execution process and makes it easier for a developer to understand and resolve any processing issues.

$instructor = (new Instructor)
    // see requests to LLM
    ->onEvent(RequestSentToLLM::class, fn($e) => dump($e))
    // see responses from LLM
    ->onEvent(ResponseReceivedFromLLM::class, fn($event) => dump($event))
    // see all events in console-friendly format
    ->wiretap(fn($event) => dump($event->toConsole()));

$instructor->respond(
    messages: "What is the population of Paris?",
    responseModel: Scalar::integer(),
);
// check your console for the details on the Instructor execution

Response Models

Instructor is able to process several types of input provided as response model, giving you more flexibility on how you interact with the library.

The signature of respond() method of Instructor states the responseModel can be either string, object or array.

Handling string $responseModel value

If string value is provided, it is used as a name of the class of the response model.

Instructor checks if the class exists and analyzes the class & properties type information & doc comments to generate a schema needed to specify LLM response constraints.

The best way to provide the name of the response model class is to use NameOfTheClass::class instead of string, making it possible for IDE to execute type checks, handle refactorings, etc.

Handling object $responseModel value

If object value is provided, it is considered an instance of the response model. Instructor checks the class of the instance, then analyzes it and its property type data to specify LLM response constraints.

Handling array $responseModel value

If array value is provided, it is considered a raw JSON Schema, therefore allowing Instructor to use it directly in LLM requests (after wrapping in appropriate context - e.g. function call).

Instructor requires information on the class of each nested object in your JSON Schema, so it can correctly deserialize the data into appropriate type.

This information is available to Instructor when you are passing $responseModel as a class name or an instance, but it is missing from raw JSON Schema.

Current design uses JSON Schema $comment field on property to overcome this. Instructor expects developer to use $comment field to provide fully qualified name of the target class to be used to deserialize property data of object or enum type.

Response model contracts

Instructor allows you to customize processing of $responseModel value also by looking at the interfaces the class or instance implements:

CanProvideJsonSchema - implement to be able to provide JSON Schema or the response model, overriding the default approach of Instructor, which is analyzing $responseModel value class information,
CanDeserializeSelf - implement to customize the way the response from LLM is deserialized from JSON into PHP object,
CanValidateSelf - implement to customize the way the deserialized object is validated,
CanTransformSelf - implement to transform the validated object into target value received by the caller (e.g. unwrap simple type from a class to a scalar value).

Additional Notes

PHP ecosystem does not (yet) have a strong equivalent of Pydantic, which is at the core of Instructor for Python.

To provide an essential functionality we needed here Instructor for PHP leverages:

base capabilities of PHP type system,
PHP reflection,
PHP DocBlock type hinting conventions,
Symfony serialization and validation capabilities

Dependencies

Instructor for PHP is compatible with PHP 8.2 or later and, due to minimal dependencies, should work with any framework of your choice.

Guzzle
Symfony components
- symfony/property-access
- symfony/property-info
- symfony/serializer
- symfony/type-info
- symfony/validator
adbario/php-dot-notation
phpdocumentor/reflection-docblock
phpstan/phpdoc-parser
vlucas/phpdotenv

Additional dependencies are required for some extras:

spatie/array-to-xml
gioni06/gpt3-tokenizer

TODOs

Async support
Documentation

Contributing

If you want to help, check out some of the issues. All contributions are welcome - code improvements, documentation, bug reports, blog posts / articles, or new cookbooks and application examples.

License

This project is licensed under the terms of the MIT License.

Support

If you have any questions or need help, please reach out to me on Twitter or GitHub.

cognesy / instructor-php

Maintainers

Details

README

What is Instructor?

Feature highlights

Core features

Flexible inputs

Customization

Sync and streaming support

Observability

Support for multiple LLMs / API providers

Other capabilities

Documentation and examples

Instructor in Other Languages

How Instructor Enhances Your Workflow

Response Model

Validation

Max Retries

Get Started

Usage

Basic example

Connecting to various LLM API providers

Structured-to-structured processing

Validation

Max Retries

Alternative ways to call Instructor

Streaming support

Partial results

Shortcuts

String as Input

Extracting Scalar Values

Extracting Enum Values

Extracting Sequences of Objects

Specifying Data Model

Type Hints

DocBlock type hints

Typed Collections / Arrays

Complex data extraction

Dynamic data schemas

Changing LLM model and options

Support for language models and API providers

Using DocBlocks as Additional Instructions for LLM

Customizing Validation

ValidationMixin

Validation Callback

Internals

Lifecycle

Receiving notification on internal events

Response Models

Handling string $responseModel value

Handling object $responseModel value

Handling array $responseModel value

Response model contracts

Additional Notes

Dependencies

TODOs

Contributing

License

Support

Contributors