Compai - Companion Chatbot

This project implements a Retrieval-Augmented Generation (RAG) Chatbot using Gemini AI, ChromaDB, and Google Palm Embeddings, integrated seamlessly with Telegram Bot. It provides a natural and engaging conversational experience with an affectionate and emotionally resonant communication style, making it feel like a caring conversational partner.

---

Presentation Slide

Features

1. Embeddings & LLM:

• Utilizes Google Palm Embeddings for efficient vector representation of text.
• Leverages Gemini Pro model (gemini-1.5-flash) for natural language generation.

2. Advanced Memorization

• Short-term memory: Stores the last five messages in a double-ended queue (deque) for maintaining conversational flow.
• Long-term memory: Extracts key facts (e.g., age, preferences, dislikes) from past conversations and stores them in a structured key-value table, allowing for more relevant and contextual responses.

3. Intelligent Conversation Retrieval

• Retrieves similar vectors from past conversations using ChromaDB, ensuring context-aware interactions.
• Merges retrieved historical context with key-value stored facts to enhance response accuracy.

4. Personalized Speaker Style Adaptation

• Analyzes writing style, tone, and emoji usage to tailor responses.
• Ensures answers mimic the user’s natural conversational flow, making the chatbot more engaging and emotionally intelligent.

5. Telegram Integration

• Allows seamless chat history uploads to automatically recognize user and speaker IDs.
• Provides real-time interaction with continuous context-awareness.

How It Works

1. Load Conversation History:

• Upload conversations file from telegram.
• Loads JSON conversations in chunks using orjson for fast processing.

2. ChromaDB Initialization:

• Initializes a ChromaDB collection and dynamically adds conversation data with vector embeddings.

3. Speaker Style Analysis:

• Analyzes the speaker's writing style, tone, and emoji usage to personalize responses.

4. Context Retrieval with ChromaDB

• When a user sends a message, the system retrieves similar past conversations using vector embeddings.
• It dynamically combines:
  • Retrieved historical context from ChromaDB.
  • Key-value structured data (e.g., user preferences).
  • The last five messages stored in short-term memory.

5. Response Generation

• Using the Gemini API, the chatbot synthesizes a response that:
  • Integrates retrieved context.
  • Matches the user’s communication style.
  • Feels natural and personalized.

---

Installation

1. Clone the repository:

git clone https://github.com/FLYIH/Compai.git
cd Compai

2. Create a virtual environment:

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install required dependencies:

pip install -r requirements.txt

API Key Setup

This project uses Google Gemini API and Telegram Bot Token for seamless integration.

• Gemini API Key: Obtain from Google AI Studio

• Telegram Bot Token: Create a bot using BotFather on Telegram.

  • Open Telegram and search for @BotFather.
  • Start a chat and use the command:

    /newbot
    

  • Follow the instructions to set a name and username for your bot.

• create a .env file in the root directory:

  GEMINI_API_KEY=your_gemini_api_key_here
  TG_BOT_TOKEN=your_telegram_bot_token_here
  

Running the Chatbot

1. Activate the virtual environment:

source venv/bin/activate

2. Run the chatbot:

python main.py

3. Interact with the Chatbot on Telegram:

• Start chatting:
  • Export chat history in JSON format.
  • Send the exported JSON file as a document attachment.
  • The chatbot will respond with context-aware and emotionally resonant replies.