Download the PHP package aisdk/openai-compatible without Composer

aisdk/openai-compatible

Shared OpenAI-compatible wire adapter for the PHP AI SDK. Reusable by any provider that speaks OpenAI-compatible Chat Completions, Responses, embedding, image generation, speech generation, or transcription APIs.

Installation

What This Package Does

This package provides the shared wire-format bridge between core portable contracts (AiSdk\*) and OpenAI-compatible provider shapes. It is used by providers like Groq, xAI, OpenRouter, and others that implement the OpenAI-compatible API.

It owns:

• Request body building (ChatRequestBuilder)
• Response parsing (ChatResponseParser)
• SSE stream parsing (ChatStreamParser)
• Responses request body building (ResponsesRequestBuilder)
• Responses request profiles (ResponsesRequestProfile)
• Responses response parsing (ResponsesResponseParser)
• Responses SSE stream parsing (ResponsesStreamParser)
• Embedding request body building (EmbeddingRequestBuilder)
• Embedding response parsing (EmbeddingResponseParser)
• Image request body building (ImageRequestBuilder)
• Image response parsing (ImageResponseParser)
• Speech request body building (SpeechRequestBuilder)
• Speech response parsing (SpeechResponseParser)
• Multipart transcription request building (TranscriptionRequestBuilder)
• Transcription response parsing (TranscriptionResponseParser)
• Message conversion (ChatMessageConverter)
• Tool conversion (ChatToolConverter)
• Usage normalization (ChatUsage)
• Finish reason mapping (MapsFinishReason)

It does not own:

• Provider authentication
• Model inventories
• Provider-specific quirks or fallback behavior

Usage

This package is consumed by provider packages, not directly by end users. A provider that speaks OpenAI-compatible chat completions uses it like this:

For Responses endpoints, use the separate Responses wire helpers rather than adapting a Chat Completions body:

For streaming, pass parsed SSE events to ResponsesStreamParser::parse() in the same way Chat Completions providers use ChatStreamParser.

ResponsesRequestProfile lets a provider declare safe endpoint differences such as its output-token field or reasoning parameter path. Unsupported portable features fail at request construction instead of being sent optimistically. Providers can still pass endpoint-specific fields through providerOptions($providerName, [...]).

For embedding endpoints:

Providers can configure the dimensions field name or omit it, and can disable the default encoding_format: float field when their endpoint uses a different wire shape. Provider-namespaced options and raw values are applied after portable fields.

For image generation endpoints:

For speech generation endpoints:

For OpenAI-compatible transcription endpoints:

Provider Integration

To build a provider on top of this package:

1. Depend on aisdk/core and aisdk/openai-compatible.
2. Create a provider class extending BaseProvider and implement protected capability hooks such as textModel() or embeddingModel().
3. Expose only Provider::model('model-id') from the public facade, delegating to the provider instance's model() method.
4. Create a text model extending BaseModel and select either the Chat Completions helpers or the Responses helpers for the provider endpoint. Both paths support text, tools, structured output, reasoning, streaming, normalized usage, and provider-namespaced options; multimodal content is used only when the provider advertises the corresponding core capability.
5. Add provider-specific auth, base URL, headers, and adapter capabilities.
6. Apply any provider-specific adaptations (e.g., structured output downgrades) after calling ChatRequestBuilder::build().

For embeddings, create a model implementing EmbeddingModelInterface, call EmbeddingRequestBuilder::build(), then parse the provider payload with EmbeddingResponseParser::parse(). For image generation, create an image model implementing ImageModelInterface, call ImageRequestBuilder::build(), then parse the provider payload with ImageResponseParser::parse(). For speech generation, create a speech model implementing SpeechModelInterface, call SpeechRequestBuilder::build(), then parse the raw audio response with SpeechResponseParser::parse(). For transcription, implement TranscriptionModelInterface, send the multipart body from TranscriptionRequestBuilder, then normalize the response with TranscriptionResponseParser. Provider packages still own authentication, endpoint paths, adapter capabilities, and public facades. Model IDs should pass through as opaque provider values instead of being maintained as a package-owned inventory.

Testing

Documentation

• PHP AI SDK documentation
• OpenAI-compatible documentation

Community

• Contributing
• Support
• For private security reports, email [email protected].