tarsier 0.2.0

Vision utilities for web interaction agents

🙈 Vision utilities for web interaction agents 🙈

🔗 Main site   •   🐦 Twitter   •   📢 Discord

Tarsier

Tried using GPT-4(V) to automate web interactions? You've probably run into issues like these:

• How do you map from an LLM's responses back to web elements?
• How do you feed a "screenshot" to a text-only LLM?
• How do you screen capture an entire page?

At Reworkd, we found ourselves reusing the same utils to solve these problems across multiple projects, so we're now open-sourcing a simple little utils library for multimodal web agents... Tarsier!

Tarsier visually tags elements on a page, allowing GPT-4V to specify by tag which element to click. Tarsier also provides OCR utils to convert a page screenshot into a whitespace-structured string that an LLM without vision can understand.

Installation

pip install tarsier

Usage

An agent using Tarsier might look like this:

from playwright.async_api import async_playwright
from tarsier import Tarsier, GoogleVisionOCRService

async def main():
    creds = {...} # Google Cloud credentials
    ocr_service = GoogleVisionOCRService(creds)
    tarsier = Tarsier(ocr_service)
    
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=False)
        page = await browser.new_page()
        await page.goto("https://news.ycombinator.com")
    
        driver = tarsier.create_driver(page)
        page_text, tag_to_xpath = await tarsier.page_to_text(driver)
        
        print(page_text) # My Text representation of the page
        print(tag_to_xpath) # Mapping of tags to x_paths

Citations

bibtex
@misc{reworkd2023tarsier,
  title        = {Tarsier},
  author       = {Rohan Pandey and Adam Watkins and Asim Shrestha and Srijan Subedi},
  year         = {2023},
  howpublished = {GitHub},
  url          = {https://github.com/reworkd/bananalyzer}
}