cognesy/instructor-php

Unified LLM API, structured data outputs with LLMs, and agent SDK - in PHP

What it solves

It eliminates the need to manually parse messy LLM outputs (like JSON or raw text) by turning them into typed, validated PHP data objects. It also solves the problem of provider lock-in by providing a unified API to interact with dozens of different LLM providers.

How it works

Instructor uses a "response model" (a plain PHP class with type-hints) to define the desired data structure. It then leverages LLM capabilities—such as native JSON mode, tool calling, or specific prompting—to force the model to return data matching that schema. If the output fails validation (using Symfony Validation), Instructor can automatically retry the request, feeding the validation errors back to the LLM so it can self-correct.

Who it’s for

PHP developers who want to integrate LLMs into their applications with type safety, structured data extraction, and the ability to switch between providers (e.g., OpenAI, Anthropic, Gemini, Ollama) without rewriting their core logic.

Highlights

  • Structured Extraction: Converts text, images, or chat sequences into typed PHP classes.
  • Automated Validation & Retries: Uses Symfony Validation to verify outputs and automatically retries failed requests with error feedback.
  • Unified Inference (Polyglot): A single API layer supporting 20+ LLM providers, including local models via Ollama.
  • Comprehensive AI Toolkit: Includes components for RAG (vector storage and retrieval), a lightweight Agent SDK for tool-using agents, and batch inference.
  • Framework Agnostic: Works standalone or via first-party integrations for Laravel and Symfony.

Related

  • Project
    567-labs/instructor

    A library that ensures LLMs return reliable, structured JSON by using Pydantic for validation, type safety, and automatic retries.

  • Project
    google/langextract

    LangExtract is a Python library for LLM‑driven, grounded extraction of structured data from any text. It chunks long documents, runs parallel LLM calls (Gemini, OpenAI, Ollama or custom providers), maps each extraction back to its exact character span, and produces an interactive HTML visualiser. Install via pip, define a prompt + few‑shot examples, call `lx.extract`, and save/visualise the results. Supports cloud, local, and batch processing, with a plugin system for new model providers.

  • Project
    mishushakov/llm-scraper

    LLM Scraper is a TypeScript library that lets you extract structured data from any webpage using large language models. It combines browser automation with schema-based prompting to turn raw web content into organized JSON, supporting multiple LLM providers and streaming output.

  • Project
    dreadnode/rigging

    A lightweight LLM framework that simplifies production code integration by providing unified model access, structured Pydantic outputs, and type-safe prompt definitions.

  • Project
    neuron-core/neuron-ai

    Neuron AI is a Composer‑installable PHP framework that lets developers build full‑featured AI agents, RAG pipelines, tool‑calling, structured output and multi‑agent workflows. It abstracts dozens of LLM providers, offers ready‑made toolkits (e.g., MySQL), an MCP connector for remote tools, and a low‑level Workflow engine for custom pipelines. Installation is a single `composer require` plus optional skill integration for Laravel/Symfony. The library targets PHP teams that want the AI agent to be the core architecture of their application.