Tezeta: A Package for Maximizing Context Window Utilization in Chatbots (Under Development)

:warning: Tezeta is currently under active development and is not yet functional. The features listed below are planned for future releases.

Tezeta is a Python package designed to optimize memory in chatbots and Language Model (LLM) requests using relevance-based vector embeddings. This tool aims to maximize the utilization of context windows, thereby improving chatbot performance by allowing the storage and retrieval of more relevant conversation history.

Supported Features

• Using vector embeddings to rank chats based on relevance with OpenAI embeddings and Pinecone
• Using ChromaDB as vector store with the default all-MiniLM-L6-v2
• Using ChromaDB as vector store
• Support for using Open Source Embedding Models locally (currently through all-MiniLM-L6-v2 with chromaDB)

Planned Features

• Chunk up and rank sections of long text in a single chat or LLM request
• Support for using the Cohere API for Embeddings

Installation

pip install tezeta

Basic Usage

First, to set the necessary environment variables in your system, you can use the following terminal commands.

For macOS/Linux:

export PINECONE_API_KEY=your_api_key
export OPENAI_API_KEY=your_api_key
export PINECONE_ENVIRONMENT=your_pinecone_environment

For Windows:

set PINECONE_API_KEY=your_api_key
set OPENAI_API_KEY=your_api_key
set PINECONE_ENVIRONMENT=your_pinecone_environment

You can use the package as follows:

import tezeta

chats = [
    {
        "role" : "user",
        "content" : "Wellness is an important part of wellbeing. How are you tackling that in your life"
    },
    {
        "role" : "user",
        "content" : "Hello there Jon, I'm a less relevant text that is trying really really hard to excluded from this test."
    },
    {
        "role" : "assistant",
        "content" : "I'm doing well, how are you?"
    }
]

tezeta.set_max_tokens(30)

print(chats)
llm_chats = tezeta.chats.fit_messages(chats)
print (llm_chats)

Documentation

Documentation will be available in the future.

License

This project is licensed under the terms of the MIT license.