LBGPT 0.1.3

Load balancer for asynchroneous requests to the APIs of OpenAI and Azure (if configured) for ChatGPT

Load Balancing ChatGPT (LBGPT)

Enhance your ChatGPT API experience with the LoadBalancing ChatGPT (LBGPT), a wrapper around OpenAI's API designed to boost performance, enable caching, and provide seamless integration with Azure's OpenAI API.

This tool significantly optimizes single request response times by asynchronously interacting with the OpenAI API and efficiently caching results. It also offers automatic retries in the event of API errors and the option to balance requests between OpenAI and Azure for an even more robust AI experience.

Proudly build by the team of Marvin Labs where we use AI to help financial analysts make better investment decisions.

Installation

You can easily install LoadBalancing ChatGPT via pip:

pip install lbgpt

Usage

Basic

Initiate asynchronous calls to the ChatGPT API using the following basic example:

import lbgpt
import asyncio

chatgpt = lbgpt.ChatGPT(api_key="YOUR_API_KEY")
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

The chat_completion_list function expects a list of dictionaries with fully-formed OpenAI ChatCompletion API requests. Refer to the OpenAI API definition for more details. You can also use the chat_completion function for single requests.

By default, LBGPT processes five requests in parallel, but you can adjust this by setting the max_concurrent_requests parameter in the constructor.

Azure

For users with an Azure account and proper OpenAI services setup, lbgpt offers an interface for Azure, similar to the OpenAI API. Here's how you can use it:

import lbgpt
import asyncio

chatgpt = lbgpt.AzureGPT(api_key="YOUR_API_KEY", azure_api_base="YOUR AZURE API BASE", azure_model_map={"OPENAI_MODEL_NAME": "MODEL NAME IN AZURE"})
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

You can use the same request definition for both OpenAI and Azure. To ensure interchangeability, map OpenAI model names to Azure model names using the azure_model_map parameter in the constructor (see https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/switching-endpoints for details).

Load Balacing OpenAI and Azure

For optimal performance and reliability, it's recommended to set up the LoadBalancedGPT or MultiLoadBalancedGPT. These classes automatically balance requests between OpenAI and Azure, and they also offer caching and automatic retries.

LoadBalancedGPT offers load-balancing just between OpenAI and Azure models, but is slightly easier to set up. By default, 75% of requests are routed to the Azure API, while 25% go to the OpenAI API. You can customize this ratio by setting the ratio_openai_to_azure parameter in the constructor, taking into account that the Azure API is considerably faster.

import lbgpt
import asyncio

chatgpt = lbgpt.LoadBalancedGPT(
    openai_api_key="YOUR_OPENAI_API_KEY",
    azure_api_key="YOUR_AZURE_API_KEY",
    azure_api_base="YOUR AZURE API BASE",
    azure_model_map={"OPENAI_MODEL_NAME": "MODEL NAME IN AZURE"})
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

MultiLoadBalancedGPT offers load-balancing between multiple OpenAI and Azure models, and offers more flexibility in terms of the load balancing inputs. In order to achieve the same load balancing as the LoadBalancedGPT, you can use the following code:

import lbgpt
import asyncio

openai_chatgpt = lbgpt.ChatGPT(api_key="YOUR_API_KEY")
azure_chatgpt = lbgpt.AzureGPT(api_key="YOUR_API_KEY", azure_api_base="YOUR AZURE API BASE", azure_model_map={"OPENAI_MODEL_NAME": "MODEL NAME IN AZURE"})


chatgpt = lbgpt.MultiLoadBalancedGPT(
    gpts=[openai_chatgpt, azure_chatgpt],
    allocation_function_weights=[0.25, 0.75],
    allocation_function='random',
)
    
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

However, the MultiLoadBalancedGPT offers more flexibility in terms of the load balancing inputs, e.g. supporting multiple Azure instances or OpenAI keys.

You can also select the allocation function max_headroom to automatically pick the API with the most available capacity. This requires you to tell the model constructors your RPM (requests per minute) and/or TPM (tokens per minute) limits.

For example, if you have an OpenAI API key with a 5,000 TPM limit and an Azure API key with a 10,000 TPM limit, you can use the following code:

import lbgpt
import asyncio

openai_chatgpt = lbgpt.ChatGPT(api_key="YOUR_API_KEY", limit_tpm=5_000)
azure_chatgpt = lbgpt.AzureGPT(api_key="YOUR_API_KEY", azure_api_base="YOUR AZURE API BASE", azure_model_map={"OPENAI_MODEL_NAME": "MODEL NAME IN AZURE"}, limit_tpm=10_000)


chatgpt = lbgpt.MultiLoadBalancedGPT(
    gpts=[openai_chatgpt, azure_chatgpt],
    allocation_function='max_headroom',
)
    
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

Caching

LBGPT implements two means of caching: basic caching and semantic cachin:

• Basic caching: Caches the exact request and response.
• Semantic caching: Caches the request and response, but allows for semantic variations in the messages in the request.

Both caching methods can be combined. If both caches are used, requests are first requested from the basic cache. Only if the request is not found in the basic cache, the semantic cache is checked. If the request is not found in the semantic cache, the request is sent to the API and the response is cached in both caches.

Basic Caching

Take advantage of request caching to avoid redundant calls:

import lbgpt
import asyncio
import diskcache

cache = diskcache.Cache("cache_dir")
chatgpt = lbgpt.ChatGPT(api_key="YOUR_API_KEY", cache=cache)
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

While LBGPT is tested only with diskcache, it should work seamlessly with any cache that implements the __getitem__ and __setitem__ methods.

Semantic Caching

Semantic caching looks for semantic variations of the request in the cache. For example, if the request message is "Hello, how are you?", the semantic cache will also return a cached response for "Hello, how are you doing?". The semantic cache uses embedding models supported by HuggingFace to determine semantic similarity.

import lbgpt
from lbgpt.semantic_cache import FaissSemanticCache
from langchain.embeddings import HuggingFaceEmbeddings

import asyncio

semantic_cache = FaissSemanticCache(
        embedding_model=HuggingFaceEmbeddings(model_name="bert-base-uncased"),
        cosine_similarity_threshold=0.95,
        path='cache_dir'
    )

chatgpt = lbgpt.ChatGPT(api_key="YOUR_API_KEY", semantic_cache=semantic_cache)
res = asyncio.run(chatgpt.chat_completion_list([ "your list of prompts" ]))

Currently, the only supported semantic caches are FAISS (via the langchain interface) and Qdrant. Please let us know if you would like to see support for other semantic caches.

In this example, all requests messages are embedded using the bert-base-uncased model from HuggingFace. The cosine similarity threshold determines how similar two messages must be to be considered semantically equivalent. The path parameter determines where the semantic cache is stored.

How to Get API Keys

To obtain your OpenAI API key, visit the official OpenAI site. For Azure API key acquisition, please refer to the official Azure documentation.