Skip to content
forked from ishan0102/vimGPT

Browse the web with GPT-4V and Vimium

License

Notifications You must be signed in to change notification settings

dakouan18/vimGPT

 
 

Repository files navigation

vimGPT

Giving multimodal models an interface to play with.

vimgpt.mov

Overview

LLMs as a way to browse the web is being explored by numerous startups and open-source projects. With this project, I was interested in seeing if we could only use GPT-4V's vision capabilities for web browsing.

The issue with this is it's hard to determine what the model wants to click on without giving it the browser DOM as text. Vimium is a Chrome extension that lets you navigate the web with only your keyboard. I thought it would be interesting to see if we could use Vimium to give the model a way to interact with the web.

Setup

Install Python requirements

pip install -r requirements.txt

Download Vimium locally (have to load the extension manually when running Playwright)

./setup.sh

Ideas

Feel free to collaborate with me on this, I have a number of ideas:

  • Use Assistant API once it's released for automatic context retrieval. The Assistant API will create a thread that we can add messages too, to keep the history of actions, but it doesn't support the Vision API yet.
  • Vimium fork for overlaying elements. A specialized version of Vimium that selectively overlays elements based on context could be useful, effectively pruning based on the user query. Might be worth testing if different sized boxes/colors help.
  • Use higher resolution images, as it seems to fail at low res. I noticed that below a certain threshold, the model wouldn't detect anything. This might be improved by using higher resolution images but that would require more tokens.
  • Fine-tune LLaVa or CogVLM to do this or Fuyu-8B. Could be faster/cheaper. CogVLM can accurately specify pixel coordinates which may be a good way to augment this.
  • Use JSON mode once it's released for Vision API. Currently the Vision API doesn't support JSON mode or function calling, so we have to rely on more primitive prompting methods.
  • Have the Vision API return general instructions, formalized by another call to the JSON mode version of the API. This is a workaround for the JSON mode issue but requires another LLM call, which is slower/more expensive.
  • Add speech-to-text with Whisper or another model to eliminate text input and make this more accessible.
  • Make this work for your own browser instead of spinning up an artificial one. I want to be able to order food with my credit card.
  • Provide the frames with and without Vimium enabled in case the model can't see what's under the yellow square.
  • Pass the Chrome accessibility tree in as input in addition to the image. This provides a layout of interactive elements that can be mapped to the Vimium bindings.
  • Have it write longer things based on the context of the page or return information to the user based on the query. Examples are replying to an email, summarizing a news article, etc. Visual question answering.
  • Make this a useful tool for blind people by adding voice mode and a key that creates an Assistant API for a given page. Something where you can "speak to an agent" about a page content in natural language.

References

About

Browse the web with GPT-4V and Vimium

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 97.7%
  • Shell 2.3%