Adapters for Large Language Models and Generative Pre-trained Transformers APIs

python -m pip install aigents

The aigents package intention is to tackle the issue of out there having a wide variety of different AI models APIs, even though most of them keep similar standards, it might be confusing or challeging for developers to merge, scale and write conssitent solutions that englobes them.

The main classes for completions or chat have Chatters on their name. For instance, aigents.GoogleChatter uses Google's AI completion models to generate chat responses. The method for calling responses is answer.

The constructors for Chatters differs a little, but on thing in common on them is the setup (str) argument: a text instructing the AI model to play some role. For example:

from aigents import GoogleChatter

setup = "You are a very experienced and successful fantasy novel writer. "
## generate your API key: https://makersuite.google.com/app/apikey
api_key = "YOUR_GOOGLE_AI_API_KEY"
temperature = 0.5  # for creativity
chatter = GoogleChatter(setup=setup, api_key=api_key)

response = chatter.answer(
    "What are the difference between soft magic system "
    "and hard magic system?"
)

The returned value is a string corresponding on the first candidate response from Gemini API, which you can access using the attribute chatter.last_response.

For OpenAI and Google chatters you can the the API key using .env file. Use the .sample.env: paste your API keys and save it as .env

By default, chatter.answer has conversation=True for every Chatter. This means that the object keeps a record of the messages (conversation), which you can access using the attribute chatter.messages.

If the messages exceeds the token window allowed for the corresponding model, the instance will try to reduce the size of the messages. If the default values of use_agent=True is set for chatter.answer, the Chatter will use another instance to summarize messages, in order to preserve most of the history and meaning of the conversation. Otherwise it will remove the oldest (but the chatter.setup, if provided) messages.

This section provides a brief overview of the classes in core.py and their usage.

Every chatter have an async version. For instance, the previous code can be written using async version of the OpenAI's chatter:

from aigents import AsyncOpenAIChatter

setup = "You are a very experienced and successful fantasy novel writer. "
## generate your API key: https://platform.openai.com/api-keys
api_key = "YOUR_OPEN_AI_API_KEY"
temperature = 0.5  # for creativity
chatter = AsyncOpenAIChatter(setup=setup, api_key=api_key)

response = await chatter.answer(
    "What are the difference between soft magic system "
    "and hard magic system?"
)

• AsyncGoogleChatter: async version of GoogleChatter;
• GoogleVision: agent for image chat (Important: gemini-pro-vision does not generate images as for this date);
• AsyncGoogleVision: async version of GoogleVision
• OpenAIChatter: chatter that uses OpenAI's API;
• BingChatter: uses the uses g4f adpter, with Bing provider;
• AsyncBingChatter: async version of BingChatter.

You can check different AI models used by aigents in module constants. For instance, for changing OpenAI chatter's model from gpt-3.5-turbo-0125 to gpt-4-0125-preview:

from aigents import OpenAIChatter
from aigents.constants import MODELS

api_key = "YOUR_OPEN_AI_API_KEY"
chatter = OpenAIChatter(setup=setup, api_key=api_key)
chatter.change_model(MODELS[4])  # gpt-4o-mini
# always checked if it is the intended model
print(chatter.model)

The Context class in aigents provides functionality for generating embeddings from a source and generating a context based on a question. This is particularly useful for contextual chat applications.

To generate embeddings from a source, you can use the generate_embeddings method. This method supports generating embeddings from a string, a pandas DataFrame, a Path to a parquet file containing embeddings, or a dictionary containing 'chunks'of text, 'n_tokens' corresponding to the number of tokens, counted using tiktoken library, and 'embeddings' which are numpy arrays of the encoded chunks as embeddings.

Here's an example of how to use the generate_embeddings method:

import asyncio
from aigents import Context

async def get_embeddings_dataframe(text):
    
    # Initialize the Context with your text
    context = Context(text=text)
    # Generate embeddings using the default model
    return await context.generate_embeddings(embedding_model='gemini')

text = ''
with open('path/to/your/text', 'r') as textfile:
    text = textfile.read()
data = asyncio.run(get_embeddings_dataframe(text))
# Print the generated embeddings
print(data)

Here we use the Google's Gemini embedding generator. Note that you'll have to set your Gemini's API key in yur environment variables. Alternatively you can pass the api key as the argument api_key in generate_embeddings method.

Currently, embeddings generation via Context is only possible using async paradigm. This type of choice is due to the underperform of when using large texts, since each chunk of text corresponds to one api call, so it's better to make asynchronous calls.

After context generation, you can extract only relevant chunks based on a given query:

import asyncio
from aigents import Context

async def get_relevant_chunks(question, text):
    embedding_model = 'gemini'
    pipeline = 'en_core_web_md'
    # Initialize the Context with your text
    context = Context(text=text)
    # Generate embeddings using the default model
    await context.generate_embeddings(embedding_model=embedding_model)
    return await context.generate_context(
        question,
        pipeline=pepiline
    )

text = ''
with open('path/to/your/text', 'r') as textfile:
    text = textfile.read()

question = 'What is this article about?'
relevant_data = asyncio.run(get_relevant_chunk(question, text))
# Print the generated embeddings
print(relevant_data)

You'll have to install the corresponding spacy pipeline. For instance, if your text is in portuguese, you'd want to install pt-core-news-sm or pt-core-news-md:

python -m spacy install pt-core-news-md

Than pass the value to the argument pipeline of the generate_context method.

Interested in contributing? Check out the contributing guidelines. Please note that this project is released with a Code of Conduct. By contributing to this project, you agree to abide by its terms.

This project is licensed under GNU_GPL_v3.0

aigents was created with cookiecutter and the py-pkgs-cookiecutter template.