README for Django LLM Interaction Wrapper

Overview

This Django application serves as a flexible, reusable wrapper designed to facilitate interactions with various large language models (LLMs) and store configurations for tools, prompts, knowledge bases, etc. It allows users to define and integrate Python-based tools into prompt templates, enabling supported LLMs to utilize these tools seamlessly.

Features

• Configurable LLM Interaction: Define and switch between multiple LLM configurations, including those from Azure, Gemini, and other platforms.
• Tool Integration: Create and manage tools that can be integrated into LLM interactions.
• Prompt Template Management: Define and manage templates for user and system prompts, ensuring dynamic and context-aware interactions.
• Knowledge Repository: Link to external knowledge sources like Azure Blob, Amazon S3, and Google Drive to enrich LLM responses.
• Chat History Tracking: Maintain and access the history of chats to analyze interactions and improve system responses over time.
• Content Referencing: Manage references to content types like PDFs and YouTube videos for use in LLM responses.

Models

1. OpenAIAssistant: Manages different LLM configurations and their associated tools. (This is additional tool to directly interact with OpenAIAssistants)
2. Tool: Represents a tool with its associated code and parameters that can be used by LLMs.
3. PromptTemplate: Manages templates for initializing conversations and structuring prompts for the LLM.
4. ChatHistory: Stores the history of interactions for analysis.
5. KnowledgeRepository: Links to external repositories for storing and retrieving additional data. (This is currently in Beta phase)
6. ContentReference: Manages references to external content utilized by the LLM. (This is currently in Beta phase)

Setup

Requirements

• Django 4.x
• Python 3.11+
• OpenAI
• Litellm

Installation

We will soon be converting this into a reusable package. Meanwhile, you can just copy this and use it as a custom app in your own repo.

LLM Configurations

Define YAML files for different LLM configurations under the directory specified in settings.LLM_CONFIGS_PATH. Example YAML configurations are provided for various LLMs such as Azure and Gemini.

Example YAMLs

name: 'gpt-4-32k-azure'
llm_config_class: 'AzureOpenAILLMConfig'
endpoint: 'your-model-endpoint'
deployment_name: 'gpt4-32k'
api_key: '<api-key>'
api_version: '2024-02-15-preview'
tools_enabled: true

name: 'gemini-pro'
llm_config_class: 'GeminiConfig'
endpoint: 'https://generativelanguage.googleapis.com/v1/models/gemini-1.0-pro:generateContent'
model_name: 'gemini-pro'
api_key: 'gemini-api-key'
tools_enabled: true

Example usage

# Example Usage for General Chat Interaction using Django LLM Wrapper

from your_app.models import ChatHistoryRepository, GenericChatDataRepository
from your_app.llm_interaction_wrapper import LLMCommunicationWrapper
from your_app.enums import ValidPromptTemplates

def handle_chat_interaction(user_id):
    # Retrieve existing chat data for the user
    chat_data = GenericChatDataRepository.get_chat_data_by_user_id(user_id)

    if chat_data.chat_history_obj is None:
        # If there is no existing chat history, initialize a new LLM wrapper
        context = {'user_id': user_id}  # Example context, can include more relevant data
        llm_wrapper = LLMCommunicationWrapper(prompt_name=ValidPromptTemplates.GENERAL_CHAT,
                                              chat_history_id=None,
                                              initialize=True,
                                              initializing_context_vars=context)
        
        # Generate an initial message to start the conversation
        message_text = "Hello! How can I assist you today?"
        message = {'role': 'assistant', 'content': message_text}

        # Add the initial message to the chat history
        llm_wrapper.chat_history_repository.add_msgs_to_chat_history([message], commit_to_db=True)

        # Update the chat data with the new chat history ID
        GenericChatDataRepository.add_chat_history_id(
            chat_data=chat_data,
            chat_history_id=llm_wrapper.get_chat_history_object().id
        )
    else:
        # If there is an existing chat history, continue using it
        llm_wrapper = LLMCommunicationWrapper(prompt_name=ValidPromptTemplates.GENERAL_CHAT,
                                              chat_history_id=chat_data.chat_history_obj.id)

    # Continue with the interaction logic here
    # e.g., processing user input, generating LLM responses, etc.

Development

• Add new LLM configurations by extending the LLMConfig class.
• Develop new tools by defining Python code and integrating them into the Tool model.
• Create prompt templates as needed for different interaction scenarios.

Coming Up

1. Basic Playground for Testing Prompt Templates
2. Support for RAG Implementation Using LlamaIndex

Contributing

Contributions to enhance functionality, improve tools, or extend support for additional LLMs are welcome. Please submit pull requests or raise issues as needed.

License

MIT License.