JarvisLite: The simplest personal assistant.

JarvisLite is a proof-of-concept that demonstrates how to build a very simple personal assistant with minimal code using a combination of OpenAI's APIs.

Features

Besides just conversing with a chatbot, it supports two more features that a personal assistant would be capable of:

• checking the weather
• manipulating categories of lists (like a TODO or grocery list)

Each of these has full natural language support: for example, you can ask whether (heh) or not you need to wear a rain jacket tomorrow.

Architecture

The architecture is relatively simple:

1. A simple state machine transitions from listening to your voice to parsing your commands to executing the requested task.
2. The transcription is interpreted as a command, which is then passed along to be interpreted as a task-specific structure.
3. The task is then executed and the result is spoken.

Visually,

speak -> Whisper -> GPT-4 -> 
  a TypeScript interface distinguishing a command -> [
    if weather:
      fetch weather at given location and time ->
        GPT4 ->
          a nice description of weather conditions
    elif lists:
      GPT4 -> how are we manipulating the list? -> do that
  ] -> speak result

Installation

python -m venv .venv
source ./.venv/bin/activate
pip install -r requirements.txt

If you get an error installing openwakeword, try dropping the line and installing it directly from the GitHub repository, instead:

pip install git+https://github.com/dscripka/openWakeWord.git

You will need an OpenAI API key. Once you have it, running the assistant is as simple as:

OPENAI_KEY="your_api_key" python -m jarvis

The activation phrase is "Hey Jarvis".