What is this?

This tool is a Python Tk interface to Gemini. Because AI Studio gets too clunky when the context gets big. This is my primary tool that I use in my research work and thought others may benefit from it. The tool also supports Llama on Vertex AI MAAS using the OpenAI API. You can easily add other LLMs as well, of course.

UI screenshot

INSTALL

Make sure Python tkinter is available, as appropriate depending on your OS and your local Python setup.

# set credentials as needed (create .env file)
# edit user_ui_model.py
# create .txt files for your System Instructions
$ pip install async-tkinter-loop google-generativeai openai
$ python3 gemini_llm_tkui.py

Configure your keys

Create .env file with:

# A) For Gemini on Google AI
# Do not set this if you want Vertex AI
GOOGLE_API_KEY=<...>

# B) For Vertex AI
GOOGLE_APPLICATION_CREDENTIALS=<...>.json
GOOGLE_VERTEX_AI_PROJECT_ID=<...>
GOOGLE_VERTEX_AI_REGION=us-central1

# C) For Llama MAAS on Vertex AI
GOOGLE_VERTEX_AI_JSON=<...>.json
GOOGLE_VERTEX_AI_MAAS_ENDPOINT=us-central1-aiplatform.googleapis.com
GOOGLE_VERTEX_AI_MAAS_PROJECT_ID=<...>
GOOGLE_VERTEX_AI_MAAS_REGION=us-central1

What can you do?

• Run LLM queries, obviously...

• Have multiple instances of the tool running (e.g. different models) and have messages cross-post more easily -- Make sure to install pyzmq -- Make sure to run the ZeroMQ broker script (gemini_llm_zmq_router.py) -- Play with the Broadcast + Q to Local buttons -- This feature will probably morph and evolve over time

• Load & save context from the tool itself

• Load context from "<> Get Code" button in Google AI Studio -- Language drop-down needs to be Python -- NB: only works with text, will fail with file attachments

• Use different configurations with different generation settings and system instructions

• Tweak generation settings from the menus and extend easily

What features are missing?

• Citations
• File attachments have had very limited testing so far
  • audio/mp3 works, I haven't tested anything else but should be as simple as mapping the mime_type
  • load/save works and de/serializes the blobs

I have no plans to currently implement these, but PRs are very welcome!