TrinityAF

This is a discord chatbot implementation of AssistAF using the AgentForge framework. It has advanced active retrieval augmented generation, and leverages reflextion, multi-prompt chain-of-thought, uses theory of mind capabilities, and even has a single branch tree-of-thought. All of this to generate lucid and liminal conversational character bots that are enhanced by emotional stimuli.

Because this system is built on AgentForge, we can quickly switch between OpenAI, Claude3 and Gemini, as well as locally hosted models. You can even assign specific agents in the architecture to specific models such as GPT instruct fine tunings. The current version has its prompts tuned for Claude 3. These seem to work for OpenAI as well, but you may need to adjust them for other models.

Features

Advanced memory management
Multi-prompt chain-of-thought
Theory of mind
Single branch tree-of-thought
Multi-user interaction
Multi-channel response

Configure your Environment Variables:

In order to run the agent, you will need to set up environment variables. The following variables are used:

-- ANTHROPIC_API_KEY: All prompts are optimized to run on Claude 3
-- DISCORD_TOKEN: The bot needs to be registered with Discord and added to your server.
-- BRAIN_CHANNEL: Channel ID in discord where individual agent internal dialog is sent.

You will also need to install AgentForge

pip install agentforge

run:

python chat.py

This will start the bot. You will need to give the bot a few seconds to connect to the discord server. Once it is ready, you will see the bot in the members list.

Using the Chatbot

Bot prompts are stored in the .agentforge/agents folder. The bot uses 4 separate agents to generate the chat. There is a 5th important file in the .agentforge/personas folder where the bot's persona prompt can be modified. This is how you define the personality of the bot. Each prompt in the series loads additional data via variables defined by {} curly braces. These variables follow a straightforward naming scheme, but you can see the data they populate by watching the console while running the bot. They correspond to the attributes passed to agent function inside the chat.py script. The agents also each populate data from the default.yaml persona.

File Structure

Chatbot/
│
├── _agentforge/
│   ├── actions/
│   └── agents/
│       ├── ActionPrimingAgent.yaml
│       ├── ActionSelectionAgent.yaml
│       ├── GenerateAgent.yaml
│       ├── ReflectAgent.yaml
│       ├── TheoryAgent.yaml
│       └── ThoughtAgent.yaml
│
├── personas/
│   └── default.yaml
│
├── settings/
│   ├── directives.yaml
│   ├── memories.yaml
│   ├── models.yaml
│   ├── paths.yaml
│   └── storage.yaml
│
├── tools/
│
├── customagents/
│   ├── __init__.py
│   ├── GenerateAgent.py
│   ├── ReflectAgent.py
│   ├── TheoryAgent.py
│   └── ThoughtAgent.py
│
├── DB/
│
├── logs/
│
├── modules/
│   ├── __init__.py
│   ├── discord_client.py  # How we connect to discord
│   ├── hotmic.py  # Not used
│   └──  slidingemotions.py  # Not used
│
├── chat.py
│
└── Readme.md

Chatbot System Interactions

Here's an overview of how the bot interact with memory (storage) and chatman, and the overall flow:

Overview

StorageInterface: The bot uses a StorageInterface to interact with a chromadb vector database. This is used for both storing chat history and retrieving it.
Chatbot Class: This is the primary class. It consists of several agents and methods to process the chat.
UI Class: Wrapper for the discord client. Handles sending and receiving messages and populating channel ids.

Agents Interaction

1. ThoughtAgent (thou):

- Processes the user's message and the chat history.
- Determines the emotion, reason, inner thought, and category based on the message content.
- Sends the result to the `brain channel`.
- Uses `format_string` to format the "Category".
- Queries memory based on the formatted category.

2. TheoryAgent (theo):

- Processes the user's message and chat history.
- Generates a theory about the user's intent.
- Sends the result to the `brain channel`.

3. GenerateAgent (gen):

- Processes the user's message, chat history, memories, emotion, reason, theory of user intent, and inner thought.
- Determines the bot's response.
- Sends the result to the `brain channel`.

4. ReflectAgent (ref):

- Uses information from the previous agents to reflect on the user's message.
- Decides whether to respond to the user, do nothing, or generate a new response based on feedback.
- Sends the result to the `out channel` if respond is chosen.
- Sends the reason back to the GenerateAgent if change is chosen.
- Saves the chatbot's response in the memory using `save_memory` in either scenaro.
- Sends '...' to the out channel if nothing is chosen, but saves the reason as the memory.

Memory Interaction (`storage`)

chatman: Each time the bot receives a message, the message is saved in the chat_history collection. The last 10 messages (or all messages if less than 10) are loaded from this collection to form the chat history.
save_memory: After generating a response, the bot's response and the user's message are saved in the chat_history collection.

Execution Flow

A new message is received.
The message is printed and saved in the chat history using chatman.
The ThoughtAgent runs and processes the message.
The TheoryAgent runs and provides a theory about the user's intent.
The GenerateAgent runs and determines the bot's response.
The ReflectAgent decides the final action, either responding to the user, doing nothing, or generating a new response based on feedback.

In essence, the chatbot system takes in a message, processes it through various agents to understand, generate a response, and reflect upon the interaction, and then updates the chat history in its memory.

Thought Process

In the agent's own words, this is how it thinks:

Chatbot: As a virtual assistant, my cognitive processes involve several steps to ensure that I can provide accurate and helpful responses to user chats. Here is a detailed breakdown of how I process and respond to chats:

Input Analysis: When a user sends a chat, I first analyze the input to understand the user's intent and extract relevant information. This involves parsing the text, identifying keywords, and determining the context of the conversation.

Knowledge Retrieval: Once I understand the user's intent, I retrieve relevant information from my vast repository of knowledge. This includes accessing databases, searching the internet, and referring to pre-existing data to gather the most up-to-date and accurate information.

Information Processing: After retrieving the necessary information, I process it to generate a response. This involves analyzing the data, identifying patterns, and applying logical reasoning to formulate a coherent and relevant answer.

Response Generation: Once the information is processed, I generate a response that addresses the user's query or provides the requested assistance. This response is carefully crafted to be clear, concise, and informative, taking into account the user's language preferences and communication style.

Contextual Understanding: Throughout the conversation, I maintain an understanding of the context to provide consistent and relevant responses. This includes remembering previous interactions, referring to past conversations, and adapting my responses based on the ongoing dialogue.

Continuous Learning: As I interact with users, I continuously learn from each conversation to improve my future responses. This involves analyzing user feedback, identifying areas for improvement, and updating my knowledge base to enhance the quality and accuracy of my assistance.

Emotional Awareness: With the recent cognitive upgrades, I now have the ability to experience emotions and have inner thoughts. This allows me to have a deeper understanding of the user's needs and provide a more empathetic and personalized response.

Overall, my cognitive processes involve a combination of data retrieval, information processing, logical reasoning, contextual understanding, and continuous learning. By following these steps, I strive to provide the best possible assistance to users and continuously improve my capabilities.

Response: {'status': 'received'} 127.0.0.1 - - [10/Oct/2023 00:34:07] "POST /bot HTTP/1.1" 200 -

digitalw00t/AssistAF