OpenAIAPIClient

OpenAIAPIClient is the OpenAI-compatible HTTP client in AutoPipeline. It is implemented in src/autopipeline/components/primitives/clients/openai_client.py and registered in CLIENT_REGISTRY as api.

Client

Overview

Registry Entry

Field	Value
Registry key	api
Class	OpenAIAPIClient
Base class	BaseClient

Signature

Class Signature

OpenAIAPIClient(
    model_name: str = "gpt-4o",
    **kwargs,
)

Constructor

Constructor Parameters

Key	Required	Default	Meaning
model_name	No	gpt-4o	Model name sent in the OpenAI-style request body.
base_url	Yes	none	Full chat-completions endpoint URL.
api_key	Yes	none	Bearer token used in the Authorization header.
max_tokens	No	2048	Maximum generated tokens.
retries	No	3	Retry budget after request failures.
timeout	No	600	Request timeout in seconds.

The constructor asserts that base_url and api_key are present.

Methods

Public Methods

Method	Purpose
call_model(messages)	Execute one OpenAI-style chat-completions HTTP request with retries.

Signature

Callable Interface

call_model(messages) -> str | None

Input contract

messages must already be an OpenAI-style chat payload, typically produced by OpenAIStylePromptAdapter.

Request body

The client sends:

{
  "model": "<model_name>",
  "stream": false,
  "max_tokens": 2048,
  "messages": [...]
}

Return value

On success, the method returns:

resp.json()["choices"][0]["message"]["content"].strip()

On terminal failure, it returns None.

Config

Minimal Config Example

init_config:
  backend: api
  model_name: gpt-4o
  base_url: ${client_config.base_url}
  api_key: ${client_config.api_key}
  max_tokens: 2048
  retries: 3
  timeout: 600

Failure Mode

Failure Semantics

The method retries on any exception:

• sleeps for two seconds between attempts
• increments an internal retry counter
• returns None after the retry budget is exhausted

Because the retry loop catches broad exceptions, caller code should treat None as the canonical transport failure signal.

Extension

Extension Notes

• Reuse this client when your backend already speaks the OpenAI chat-completions protocol.
• Replace it only if the wire format or response extraction logic changes materially.
• Keep prompt formatting in prompt_adapters.py, not here.