any-llm-client

A unified and lightweight asynchronous Python API for communicating with LLMs.

Supports multiple providers, including OpenAI Chat Completions API (and any OpenAI-compatible API, such as Ollama and vLLM) and YandexGPT API.

How To Use

Before starting using any-llm-client, make sure you have it installed:

uv add any-llm-client
poetry add any-llm-client

Response API

Here's a full example that uses Ollama and Qwen2.5-Coder:

import asyncio

import any_llm_client


config = any_llm_client.OpenAIConfig(url="http://127.0.0.1:11434/v1/chat/completions", model_name="qwen2.5-coder:1.5b")


async def main() -> None:
    async with any_llm_client.get_client(config) as client:
        print(await client.request_llm_message("Кек, чо как вообще на нарах?"))


asyncio.run(main())

To use YandexGPT, replace the config:

config = any_llm_client.YandexGPTConfig(
    auth_header=os.environ["YANDEX_AUTH_HEADER"], folder_id=os.environ["YANDEX_FOLDER_ID"], model_name="yandexgpt"
)

Streaming API

LLMs often take long time to respond fully. Here's an example of streaming API usage:

import asyncio

import any_llm_client


config = any_llm_client.OpenAIConfig(url="http://127.0.0.1:11434/v1/chat/completions", model_name="qwen2.5-coder:1.5b")


async def main() -> None:
    async with (
        any_llm_client.get_client(config) as client,
        client.stream_llm_message_chunks("Кек, чо как вообще на нарах?") as message_chunks,
    ):
        async for chunk in message_chunks:
            print(chunk, end="", flush=True)


asyncio.run(main())

Passing chat history and temperature

You can pass list of messages instead of str as the first argument, and set temperature:

async with (
    any_llm_client.get_client(config) as client,
    client.stream_llm_message_chunks(
        messages=[
            any_llm_client.SystemMessage("Ты — опытный ассистент"),
            any_llm_client.UserMessage("Кек, чо как вообще на нарах?"),
        ],
        temperature=1.0,
    ) as message_chunks,
):
    ...

Other

Mock client

You can use a mock client for testing:

config = any_llm_client.MockLLMConfig(
    response_message=...,
    stream_messages=["Hi!"],
)

async with any_llm_client.get_client(config, ...) as client:
    ...

Configuration with environment variables

Credentials

Instead of passing credentials directly, you can set corresponding environment variables:

• OpenAI: ANY_LLM_CLIENT_OPENAI_AUTH_TOKEN,
• YandexGPT: ANY_LLM_CLIENT_YANDEXGPT_AUTH_HEADER, ANY_LLM_CLIENT_YANDEXGPT_FOLDER_ID.

LLM model config (with pydantic-settings)

import os

import pydantic_settings

import any_llm_client


class Settings(pydantic_settings.BaseSettings):
    llm_model: any_llm_client.AnyLLMConfig


os.environ["LLM_MODEL"] = """{
    "api_type": "openai",
    "url": "http://127.0.0.1:11434/v1/chat/completions",
    "model_name": "qwen2.5-coder:1.5b"
}"""
settings = Settings()

async with any_llm_client.get_client(settings.llm_model, ...) as client:
    ...

Combining with environment variables from previous section, you can keep LLM model configuration and secrets separate.

Using clients directly

The recommended way to get LLM client is to call any_llm_client.get_client(). This way you can easily swap LLM models. If you prefer, you can use any_llm_client.OpenAIClient or any_llm_client.YandexGPTClient directly:

config = any_llm_client.OpenAIConfig(
    url=pydantic.HttpUrl("https://api.openai.com/v1/chat/completions"),
    auth_token=os.environ["OPENAI_API_KEY"],
    model_name="gpt-4o-mini",
)

async with any_llm_client.OpenAIClient(config, ...) as client:
    ...

Errors

any_llm_client.LLMClient.request_llm_message() and any_llm_client.LLMClient.stream_llm_message_chunks() will raise any_llm_client.LLMError or any_llm_client.OutOfTokensOrSymbolsError when the LLM API responds with a failed HTTP status.

Timeouts, proxy & other HTTP settings

Pass custom HTTPX kwargs to any_llm_client.get_client():

import httpx

import any_llm_client


async with any_llm_client.get_client(
    ...,
    mounts={"https://api.openai.com": httpx.AsyncHTTPTransport(proxy="http://localhost:8030")},
    timeout=httpx.Timeout(None, connect=5.0),
) as client:
    ...

Default timeout is httpx.Timeout(None, connect=5.0) (5 seconds on connect, unlimited on read, write or pool).

Retries

By default, requests are retried 3 times on HTTP status errors. You can change the retry behaviour by supplying request_retry parameter:

async with any_llm_client.get_client(..., request_retry=any_llm_client.RequestRetryConfig(attempts=5, ...)) as client:
    ...

Passing extra data to LLM

await client.request_llm_message("Кек, чо как вообще на нарах?", extra={"best_of": 3})