Download the PHP package xp-forge/openai without Composer

OpenAI APIs for XP

This library implements OpenAI APIs with a low-level abstraction approach, supporting their REST and realtime APIs, request and response streaming, function calling and TikToken encoding.

Quick start

Using the REST API, see https://platform.openai.com/docs/api-reference/making-requests

Streaming

The REST API can use server-sent events to stream responses, see https://platform.openai.com/docs/api-reference/streaming

To access the result object, check for the response.completed event type and use its value. It contains the outuputs as well as model, filter results and usage information.

TikToken

Encodes text to tokens. Download the vocabularies cl100k_base (used for GPT-3.5 and GPT-4.0) and o200k_base (used for Omni and O1) first!

Instead of encode(), you can use count() to count the number of tokens.

Embeddings

To create an embedding for a given text, use https://platform.openai.com/docs/guides/embeddings/what-are-embeddings

Text to speech

To stream generate audio, use the API's transmit() method, which sends the given payload and returns the response. See https://platform.openai.com/docs/guides/text-to-speech/overview

Speech to text

To convert audio into text, upload files via the API's open() method, which returns an Upload instance. See https://platform.openai.com/docs/guides/speech-to-text/overview

You can also stream uploads from InputStreams as follows:

Tracing the calls

REST API calls can be traced with the logging library:

Tool calls

There are two types of tools: Built-ins like file_search and code_interpreter (available in the assistants API) as well as custom functions, see https://platform.openai.com/docs/guides/function-calling

Defining functions

Custom functions map to instance methods in a class:

The Param annnotation may define a description and a JSON schema type:

• #[Param('The name of the city')] $name
• #[Param(type: ['type' => 'string', 'enum' => ['C', 'F']])] $unit

Passing custom functions

Custom functions are registered in a Functions instance and passed via tools inside the payload.

Invoking custom functions

If tool calls are requested by the LLM, invoke them and return to next completion cycle. See https://platform.openai.com/docs/guides/function-calling/configuring-parallel-function-calling

Passing context

Functions can be passed a context as follows by annotating parameters with the Context annotation:

Azure OpenAI

These endpoints differ slightly in how they are invoked, which is handled by the AzureAI implementation. See https://learn.microsoft.com/en-us/azure/ai-services/openai/overview

Distributing requests

The Distributed endpoint allows to distribute requests over multiple endpoints. The ByRemainingRequests class uses the x-ratelimit-remaining-requests header to determine the target. See https://platform.openai.com/docs/guides/rate-limits

For more complex load balancing, have a look at this blog article using Azure API management

Realtime API

The realtime API allows streaming audio and/or text to and from language models, see https://platform.openai.com/docs/guides/realtime

For Azure AI, the setup code is slightly different:

Completions API

To use the legacy (but industry standard) chat completions API, see https://platform.openai.com/docs/quickstart?api-mode=chat:

The result object is computed from the streamed deltas and can be retrieved by accessing $flow->result().

See also

• https://github.com/openai/tiktoken/
• https://github.com/openai/openai-python
• https://github.com/openai/openai-node
• https://github.com/Azure-Samples/azure-openai-reverse-proxy
• https://www.youtube.com/watch?v=i-oHvHejdsc - GPT Function calling in a nutshell