👩‍🏫 adrienbrault/instructrice

Typing LLM completions

Best in class LLMs are able to output JSON following a schema you provide, usually JSON-Schema. This significantly expands the ways you can leverage LLMs in your application!

Think of the input as:

• A context, anything that is or can be converted to text, like emails/pdfs/html/xlsx
• A schema, "Here is the form you need to fill to complete your task"
• An optional prompt, giving a specific task, rules, etc

And the output/outcome is whichever structure best matches your use case and domain.

The python instructor cookbook has interesting examples.

Introduction

Instructrice is a PHP library that simplifies working with structured output from LLMs in a type-safe manner.

Features:

• Flexible schema options:
  • Classes using api-platform/json-schema
  • Dynamically generated types PSL\Type
  • Or a JSON-Schema array generated by a third party library, or in plain PHP
• symfony/serializer integration to deserialize LLMs outputs
• Streaming first:
  • As a developer you can be more productive with faster feedback loops than waiting for outputs to complete. This also makes slower local models more usable.
  • You can provide a much better and snappier UX to your users.
  • The headaches of parsing incomplete JSON are handled for you.
• A set of pre-configured LLMs with the best available settings. Set your API keys and switch between different providers and models without having to think about the model name, json mode, function calling, etc.

A Symfony Bundle is also available.

Installation and Usage

$ composer require adrienbrault/instructrice:@dev

use AdrienBrault\Instructrice\InstructriceFactory;
use AdrienBrault\Instructrice\LLM\Provider\Ollama;
use AdrienBrault\Instructrice\LLM\Provider\OpenAi;
use AdrienBrault\Instructrice\LLM\Provider\Anthropic;

$instructrice = InstructriceFactory::create(
    defaultLlm: Ollama::HERMES2THETA_LLAMA3_8B,
    apiKeys: [ // Unless you inject keys here, api keys will be fetched from environment variables
        OpenAi::class => $openAiApiKey,
        Anthropic::class => $anthropicApiKey,
    ],
);

List of object

use AdrienBrault\Instructrice\Attribute\Prompt;

class Character
{
    // The prompt annotation lets you add instructions specific to a property
    #[Prompt('Just the first name.')]
    public string $name;
    public ?string $rank = null;
}

$characters = $instructrice->getList(
    Character::class,
    'Colonel Jack O\'Neil walks into a bar and meets Major Samanta Carter. They call Teal\'c to join them.',
);

/*
dump($characters);
array:3 [
  0 => Character^ {
    +name: "Jack"
    +rank: "Colonel"
  }
  1 => Character^ {
    +name: "Samanta"
    +rank: "Major"
  }
  2 => Character^ {
    +name: "Teal'c"
    +rank: null
  }
]
*/

Object

$character = $instructrice->get(
    type: Character::class,
    context: 'Colonel Jack O\'Neil.',
);

/*
dump($character);
Character^ {
  +name: "Jack"
  +rank: "Colonel"
}
*/

Dynamic Schema

$label = $instructrice->get(
    type: [
        'type' => 'string',
        'enum' => ['positive', 'neutral', 'negative'],
    ],
    context: 'Amazing great cool nice',
    prompt: 'Sentiment analysis',
);

/*
dump($label);
"positive"
*/

You can also use third party json schema libraries like goldspecdigital/oooas to generate the schema:

• examples/oooas.php

CleanShot.2024-04-18.at.14.11.39.mp4

Supported providers

Provider	Environment Variables	Enum	API Key Creation URL
Ollama	OLLAMA_HOST	Ollama
OpenAI	OPENAI_API_KEY	OpenAi	API Key Management
Anthropic	ANTHROPIC_API_KEY	Anthropic	API Key Management
Mistral	MISTRAL_API_KEY	Mistral	API Key Management
Fireworks AI	FIREWORKS_API_KEY	Fireworks	API Key Management
Groq	GROQ_API_KEY	Groq	API Key Management
Together AI	TOGETHER_API_KEY	Together	API Key Management
Deepinfra	DEEPINFRA_API_KEY	Deepinfra	API Key Management
Perplexity	PERPLEXITY_API_KEY	Perplexity	API Key Management
Anyscale	ANYSCALE_API_KEY	Anyscale	API Key Management
OctoAI	OCTOAI_API_KEY	OctoAI	API Key Management

The supported providers are Enums, which you can pass to the llm argument of InstructriceFactory::create:

use AdrienBrault\Instructrice\InstructriceFactory;
use AdrienBrault\Instructrice\LLM\Provider\OpenAi;

$instructrice->get(
    ...,
    llm: OpenAi::GPT_4T, // API Key will be fetched from the OPENAI_API_KEY environment variable
);

Supported models

Strategy	📄 Text	🧩 JSON	🚀 Function

Commercial usage 💼	✅ Yes	⚠️ Yes, but	❌ Nope

Open Weights

Foundation

	💼	ctx	Ollama	Mistral	Fireworks	Groq	Together	DeepInfra	Perplexity	Anyscale	OctoAI
Mistral 7B	✅	32k	🧩	🧩 68/s			📄 98/s		📄 88/s !ctx=16k!	🧩	🧩
Mixtral 8x7B	✅	32k	🧩	🧩 44/s	🧩 237/s	🚀 560/s	🚀 99/s		📄 119/s !ctx=16k!	🧩	🧩
Mixtral 8x22B	✅	65k	🧩	🧩 77/s	🧩 77/s		📄 52/s	🧩 40/s	📄 62/s !ctx=16k!	🧩	🧩
Phi-3-Mini-4K	✅	4k	🧩
Phi-3-Mini-128K	✅	128k	🧩
Phi-3-Medium-4K	✅	4k	🧩
Phi-3-Medium-128K	✅	128k	🧩
Qwen2 0.5B	✅	32k	🧩
Qwen2 1.5B	✅	32k	🧩
Qwen2 7B	✅	128k	🧩
Llama3 8B	⚠️	8k	📄		🧩 280/s	🚀 800/s	📄 194/s	🧩 133/s	📄 121/s	🧩	🧩
Llama3 70B	⚠️	8k	🧩		🧩 116/s	🚀 270/s	📄 105/s	🧩 26/s	📄 42/s	🧩	🧩
Llama 3.1 8B	[⚠️][llama31_license]	128k	🧩		🧩	📄	📄	🧩			📄
Llama 3.1 70B	[⚠️][llama31_license]	128k	🧩		🧩	📄	📄	🧩			📄
Llama 3.1 405B	[⚠️][llama31_license]	128k	🧩		🧩	📄	📄				📄
Gemma 7B	⚠️	8k				🚀 800/s	📄 118/s	🧩 64/s		🧩
DBRX	⚠️	32k			🧩 50/s		📄 72/s	🧩
Qwen2 72B	⚠️	128k	🧩
Qwen1.5 32B	⚠️	32k					📄				🧩
Command R	❌	128k	📄
Command R+	❌	128k	📄

Throughputs from https://artificialanalysis.ai/leaderboards/providers .

Fine Tune

	💼	ctx	Base	Ollama	Fireworks	Together	DeepInfra	OctoAI
Hermes 2 Pro Mistral 7B	✅		Mistral 7B	🧩	🧩			🧩
FireFunction V1	✅		Mixtral 8x7B		🚀
WizardLM 2 7B	✅		Mistral 7B				🧩
WizardLM 2 8x22B	✅		Mixtral 8x7B			📄	🧩	🧩
Capybara 34B	✅	200k	Yi 34B		🧩
Hermes 2 Pro Llama3 8B	⚠️		Llama3 8B	📄
Hermes 2 Theta Llama3 8B	⚠️		Llama3 8B	📄
Dolphin 2.9	⚠️	8k	Llama3 8B	🧩		📄	🧩

Proprietary

Provider	Model	ctx
Mistral	Large	32k	✅ 26/s
OpenAI	GPT-4o	128k	🚀 83/s
OpenAI	GPT-4o mini	128k	🚀 140/s
OpenAI	GPT-4 Turbo	128k	🚀 28/s
OpenAI	GPT-3.5 Turbo	16k	🚀 72/s
Anthropic	Claude 3 Haiku	200k	📄 88/s
Anthropic	Claude 3 Sonnet	200k	📄 59/s
Anthropic	Claude 3 Opus	200k	📄 26/s
Google	Gemini 1.5 Flash	1000k	🧩 136/s
Google	Gemini 1.5 Pro	1000k	🧩 57/s
Perplexity	Sonar Small Chat	16k	📄
Perplexity	Sonar Small Online	12k	📄
Perplexity	Sonar Medium Chat	16k	📄
Perplexity	Sonar Medium Online	12k	📄

Throughputs from https://artificialanalysis.ai/leaderboards/providers .

Automate updating these tables by scraping https://artificialanalysis.ai , along with chatboard arena elo.? Would be a good use case / showcase of this library/cli?

Custom Models

Ollama

If you want to use an Ollama model that is not available in the enum, you can use the Ollama::create static method:

use AdrienBrault\Instructrice\LLM\LLMConfig;
use AdrienBrault\Instructrice\LLM\Cost;
use AdrienBrault\Instructrice\LLM\OpenAiJsonStrategy;
use AdrienBrault\Instructrice\LLM\Provider\Ollama;

$instructrice->get(
    ...,
    llm: Ollama::create(
        'codestral:22b-v0.1-q5_K_M', // check its license first!
        32000,
    ),
);

OpenAI

You can also use any OpenAI compatible api by passing an LLMConfig:

use AdrienBrault\Instructrice\LLM\LLMConfig;
use AdrienBrault\Instructrice\LLM\Cost;
use AdrienBrault\Instructrice\LLM\OpenAiJsonStrategy;

$instructrice->get(
    ...,
    llm: new LLMConfig(
        uri: 'https://api.together.xyz/v1/chat/completions',
        model: 'meta-llama/Llama-3-70b-chat-hf',
        contextWindow: 8000,
        label: 'Llama 3 70B',
        provider: 'Together',
        cost: Cost::create(0.9),
        strategy: OpenAiJsonStrategy::JSON,
        headers: [
            'Authorization' => 'Bearer ' . $apiKey,
        ]
    ),
);

DSN

You may configure the LLM using a DSN:

• the scheme is the provider: openai, openai-http, anthropic, google
• the password is the api key
• the host, port and path are the api endpoints without the scheme
• the query string:
  • model is the model name
  • context is the context window
  • strategy is the strategy to use:
    • json for json mode with the schema in the prompt only
    • json_with_schema for json mode with probably the completion perfectly constrained to the schema
    • tool_any
    • tool_auto
    • tool_function

Examples:

use AdrienBrault\Instructrice\InstructriceFactory;

$instructrice = InstructriceFactory::create(
    defaultLlm: 'openai://:[email protected]/v1/chat/completions?model=gpt-3.5-turbo&strategy=tool_auto&context=16000'
);

$instructrice->get(
    ...,
    llm: 'openai-http://localhost:11434?model=adrienbrault/nous-hermes2theta-llama3-8b&strategy=json&context=8000'
);

$instructrice->get(
    ...,
    llm: 'openai://:[email protected]/inference/v1/chat/completions?model=accounts/fireworks/models/llama-v3-70b-instruct&context=8000&strategy=json_with_schema'
);

$instructrice->get(
    ...,
    llm: 'google://:[email protected]/v1beta/models?model=gemini-1.5-flash&context=1000000'
);

$instructrice->get(
    ...,
    llm: 'anthropic://:[email protected]?model=claude-3-haiku-20240307&context=200000'
);

LLMInterface

You may also implement LLMInterface.

Acknowledgements

Obviously inspired by instructor-php and instructor.

How is it different from instructor php?

Both libraries essentially do the same thing:

• Automatic schema generation from classes
• Multiple LLM/Providers abstraction/support
• Many strategies to extract data: function calling, json mode, etc
• Automatic deserialization/hydration
• Maybe validation/retries later for this lib.

However, instructice differs with:

• Streaming first.
• Preconfigured provider+llms, to not have to worry about:
  • Json mode, function calling, etc
  • The best prompt format to use
  • Your options for local models
  • Whether streaming works. For example, groq can only do streaming without json-mode/function calling.
• PSR-3 logging
• Guzzle+symfony/http-client support
• No messages. You just pass context, prompt.
  • I am hoping that this choice enables cool things later like supporting few-shots examples, evals, etc
• More flexible schema options
• Higher level abstraction. You aren't able to provide a list of messages, while it is possible with instructor-php.

Notes/Ideas

Things to look into:

• Unstructured
• Llama Parse
• EMLs
• jina-ai/reader -> This is awesome, $client->request('GET', 'https://r.jina.ai/' . $url)
• firecrawl

DSPy is very interesting. There are great ideas to be inspired by.

Ideally this library is good to prototype with, but can support more advanced extraction workflows with few shot examples, some sort of eval system, generating samples/output like DSPy, etc

Would be cool to have a CLI, that accepts a FQCN and a context.

instructrice get "App\Entity\Customer" "$(cat some_email_body.md)" 

Autosave all input/schema/output in sqlite db. Like llm? Leverage that to test examples, add few shots, evals?