Skip to content

Latest commit

 

History

History
162 lines (122 loc) · 5.64 KB

readme.md

File metadata and controls

162 lines (122 loc) · 5.64 KB
Raw

German AI Voice Assistant with RAG

This project implements a German-speaking AI assistant using Retrieval-Augmented Generation (RAG) with LangChain, Groq API, and Chroma for vector storage.

Table of Contents

Features

  • Uses Groq API with the llama-3.1-70b-versatile
  • Implements LangChain for data management and processing
  • Utilizes Chroma for chat history storage
  • Employs nomic-embed-text:latest from Ollama for embeddings
  • Implements text chunking for efficient processing
  • Allows user interaction through text or speech input
  • Manages configuration data using dotenv

Prerequisites

  • Python 3.8+ (I use 3.11)
  • Groq API access
  • Chroma
  • Ollama running locally

Installation

  1. Clone the repository:
git clone https://github.com/kabelklaus/AI-Voice-Assistant.git
cd AI-Voice-Assistant
  1. Install the required packages:
pip install -r requirements.txt
  1. Rename .env_example to .env and fill in your API keys and other configuration details:
mv .env_example .env
  1. Edit the .env file with your specific credentials and settings.

MacOS Users: Installing PyAudio

If you're using a Mac and need to install PyAudio, follow these steps:

  1. Install Xcode from the App Store and restart your computer.
  2. Run the following commands in sequence:
xcode-select --install
brew remove portaudio
brew install portaudio
pip3 install pyaudio

Note: Xcode command line tools are required for some installations. Homebrew requires Xcode, so you can also just run:

xcode-select --install

Downloading Ollama Embedding Model

This project uses the nomic-embed-text:latest model from Ollama for embeddings. To download this model:

  1. Ensure Ollama is installed on your system. If not, follow the installation instructions at Ollama's official website.
  2. Open a terminal or command prompt.
  3. Run the following command to download the model:
ollama pull nomic-embed-text:latest

This command will download and install the latest version of the nomic-embed-text model. 4. Wait for the download to complete. The model size is approximately 670MB, so it may take a few minutes depending on your internet connection.

Once the model is downloaded, Ollama will automatically use it for generating embeddings in this project.

Usage

  1. Ensure Ollama is running locally for embeddings.

  2. Run the main script:

python main.py
  1. Follow the prompts to interact with the AI assistant. You can choose between text or speech input.

Adding New Skills

To extend the assistant's capabilities, you can add new skills:

  1. Create a new Python file in the skills directory (e.g., new_skill.py).
  2. Implement the skill's functionality.
  3. Import and integrate the new skill in main.py.
  4. Add the skill to the response_prompt.py file to make the LLM aware of it. Include an example of how to use the skill to help the LLM better understand and utilize it.

Example addition to response_prompt.py:

If the user asks about the weather, respond with:
FUNCTION_CALL: get_weather(location)
For example:
- If the user asks "Wie ist das Wetter in Berlin?", respond with:
FUNCTION_CALL: get_weather("Berlin")

This will ensure that the LLM knows how to use the new skill and can incorporate it into its responses.

Roadmap

  • We are considering switching from AstraDB to ChromaDB for vector storage. For more privacy.
  • More skills will be added to enhance the assistant's capabilities.
  • Regular database cleaning
    • Implement a cleaning strategy while preserving entries with "info_type" in metadata
    • Cleaning methods to consider:
      • Time-based cleaning
      • Redundancy removal
      • Context cleaning
      • Frequency-based cleaning
      • Size limitation

Database Cleaning Strategy

Our database cleaning strategy will focus on maintaining relevant information while optimizing storage and performance. Here's a breakdown of our approach:

  1. Time-based Cleaning:

    • Remove older entries not marked as user information (no "info_type" in metadata)
    • E.g., delete entries older than 30 days that aren't saved user information

  2. Redundancy Removal:

    • Identify very similar entries (high similarity in embeddings)
    • Keep only the most recent entry
    • Exclude entries with "info_type" in metadata from this process

  3. Context Cleaning:

    • Remove context information from past conversations that's no longer relevant
    • Always retain the most important or frequently accessed information

  4. Frequency-based Cleaning:

    • Remove entries that are rarely or never retrieved, except those with an "info_type"

  5. Size Limitation:

    • Implement a maximum number of entries or maximum database size
    • When the limit is reached, remove the oldest entries (excluding "info_type" entries)

This strategy will help us maintain a clean, efficient database while preserving crucial user information.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.