openai-ratelimiter

openai-ratelimiter is a simple and efficient rate limiter for the OpenAI API. It is designed to help prevent the API rate limit from being reached when using the OpenAI library. Currently, it supports only Redis as the caching service.

Note: This package has been tested with Python 3.11.4.

Installation

To install the openai-ratelimiter library, use pip:

pip install openai-ratelimiter

Redis Setup

This library uses Redis for caching. If you don't have a Redis server setup, you can pull the Redis Docker image and run a container as follows:

# Pull the Redis image
docker pull redis

# Run the Redis container
docker run --name some-redis -p 6379:6379 -d redis

This will set up a Redis server accessible at localhost on port 6379.

Usage

The library provides two classes, ChatCompletionLimiter and TextCompletionLimiter, for limiting the rate of API calls.

ChatCompletionLimiter

from openai_ratelimiter import ChatCompletionLimiter
import openai

openai.api_key = "{your API key}"
model_name = "gpt-3.5-turbo-16k"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the capital of Morocco."},
]
max_tokens = 200
chatlimiter = ChatCompletionLimiter(
    model_name=model_name,
    RPM=3_000,
    TPM=250_000,
    redis_host="localhost",
    redis_port=6379,
)
with chatlimiter.limit(messages=messages, max_tokens=max_tokens):
    response = openai.ChatCompletion.create(
        model=model_name, messages=messages, max_tokens=max_tokens
    )
    ...

TextCompletionLimiter

from openai_ratelimiter import TextCompletionLimiter
import openai

openai.api_key = "{your API key}"
model_name = "text-davinci-003"
prompt = "What is the capital of Morocco."
max_tokens = 200
textlimiter = TextCompletionLimiter(
    model_name=model_name,
    RPM=3_000,
    TPM=250_000,
    redis_host="localhost",
    redis_port=6379,
)
with textlimiter.limit(prompt=prompt, max_tokens=max_tokens):
    response = openai.Completion.create(
        model=model_name, prompt=prompt, max_tokens=max_tokens
    )
    ...

Note: The rate limits (RPM and TPM) and the Redis host and port provided in the examples are not universal and should be tailored to your specific use case. Please adjust these parameters in accordance with the selected model and your account's rate limits. To find your specific rate limits, please refer to your OpenAI account settings at OpenAI Rate Limits.

Available Methods for Limiter Classes

The following methods are available in all Limiter classes including ChatCompletionLimiter, AsyncChatCompletionLimiter, BaseAPILimiterRedis, and AsyncBaseAPILimiterRedis:

`clear_locks()`

The clear_locks() method is used to remove all locks that are currently associated with the model. This method is useful when you need to manually reset the lock state.

Usage:

limiter = AsyncChatCompletionLimiter(
        model_name=model_name,
        RPM=3_000,  
        TPM=180_000,  
        redis_host="localhost",
        redis_port=6379,
    )
success = limiter.clear_locks()

In this example, success will be True if the locks were cleared successfully, otherwise it will be False.

`is_locked(messages: List[Dict[str, str]], max_tokens: int)`

The is_locked() method is used to check if the request would be locked given the specified messages and max tokens.

Usage:

limiter = AsyncChatCompletionLimiter(
        model_name=model_name,
        RPM=3_000,  
        TPM=180_000,  
        redis_host="localhost",
        redis_port=6379,
    )
messages = [{"role": "system", "content": "You are a helpful assistant."}]
max_tokens = 200
lock_state = limiter.is_locked(messages=messages, max_tokens=max_tokens)

In this example, lock_state will be True if the request would be locked given the messages and max tokens, otherwise it will be False.

This should provide users with a clear understanding of how to use the clear_locks and is_locked methods with any of the Limiter classes.

Asynchronous Programming Support

This library also provides support for asynchronous programming with two classes: AsyncChatCompletionLimiter and AsyncTextCompletionLimiter. You can import these classes as follows:

from openai_ratelimiter.asyncio import AsyncChatCompletionLimiter, AsyncTextCompletionLimiter

The methods for these classes are the same as their synchronous counterparts. However, the context managers for these classes are asynchronous and must run within an async function.

Here are some examples of how to use these classes:

AsyncChatCompletionLimiter

import asyncio
import openai
from openai_ratelimiter.asyncio import AsyncChatCompletionLimiter

openai.api_key = "{Openai API key}"
model_name = "gpt-3.5-turbo-16k"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the capital of Morocco."},
]
max_tokens = 175
chatlimiter = AsyncChatCompletionLimiter(
    model_name=model_name,
    RPM=3_500,
    TPM=180_000,
    redis_host="localhost",
    redis_port=6379,
)

async def send_request():
    async with chatlimiter.limit(messages=messages, max_tokens=max_tokens):
        response = await openai.ChatCompletion.acreate(
            model=model_name, messages=messages, max_tokens=max_tokens
        )
        # process response here...

async def main():
    async with asyncio.TaskGroup() as tg:
        for _ in range(100):
            tg.create_task(send_request())

asyncio.run(main())

AsyncTextCompletionLimiter

import asyncio
import openai
from openai_ratelimiter.asyncio import AsyncTextCompletionLimiter

openai.api_key = "{OpenAI API key}"
model_name = "text-davinci-003"
prompt = "What is the capital of Morocco."
max_tokens = 200

textlimiter = AsyncTextCompletionLimiter(
    model_name=model_name,
    RPM=3_500,
    TPM=180_000,
    redis_host="localhost",
    redis_port=6379,
)

async def send_request(_):
    async with textlimiter.limit(prompt=prompt, max_tokens=max_tokens):
        print(_)
        response = await openai.Completion.acreate(
            model=model_name, prompt=prompt, max_tokens=max_tokens
        )
        # process response here...

async def main():
    async with asyncio.TaskGroup() as tg:
        for _ in range(100):
            tg.create_task(send_request(_))

asyncio.run(main())

Future Plans

In-memory caching
Limiting for embeddings
Limiting for DALL·E image model
Implementing more functions that provide information about the current state
Implement limiting for the organization level.

Contributing

Contributions are welcome! Please feel free to submit a pull request or open an issue on the GitHub repository. Before contributing, make sure to read through any contributing guidelines and adhere to the code of conduct.

Author

This library is created and maintained by Youssef Benhammouda. If you have any questions or feedback, you can reach out to him through the GitHub repository.

jucor/openai-ratelimiter