🙈 Vision utilities for web interaction agents 🙈

Tarsier

If you've tried using an LLM to automate web interactions, you've probably run into questions like:

How should you feed the webpage to an LLM? (e.g. HTML, Accessibility Tree, Screenshot)
How do you map LLM responses back to web elements?
How can you inform a text-only LLM about the page's visual structure?

At Reworkd, we iterated on all these problems across tens of thousands of real web tasks to build a powerful perception system for web agents... Tarsier! In the video below, we use Tarsier to provide webpage perception for a minimalistic GPT-4 LangChain web agent.

tarsier.mp4

How does it work?

Tarsier visually tags interactable elements on a page via brackets + an ID e.g. [23]. In doing this, we provide a mapping between elements and IDs for an LLM to take actions upon (e.g. CLICK [23]). We define interactable elements as buttons, links, or input fields that are visible on the page; Tarsier can also tag all textual elements if you pass tag_text_elements=True.

Furthermore, we've developed an OCR algorithm to convert a page screenshot into a whitespace-structured string (almost like ASCII art) that an LLM even without vision can understand. Since current vision-language models still lack fine-grained representations needed for web interaction tasks, this is critical. On our internal benchmarks, unimodal GPT-4 + Tarsier-Text beats GPT-4V + Tarsier-Screenshot by 10-20%!

Tagged Screenshot	Tagged Text Representation

Installation

pip install tarsier

Usage

Visit our cookbook for agent examples using Tarsier:

An autonomous LangChain web agent 🦜⛓️
An autonomous LlamaIndex web agent 🦙

Otherwise, basic Tarsier usage might look like the following:

import asyncio

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

def load_google_cloud_credentials(json_file_path):
    with open(json_file_path) as f:
        credentials = json.load(f)
    return credentials

async def main():
    # To create the service account key, follow the instructions on this SO answer https://stackoverflow.com/a/46290808/1780891
    google_cloud_credentials = load_google_cloud_credentials('./google_service_acc_key.json')

    ocr_service = GoogleVisionOCRService(google_cloud_credentials)
    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")

        page_text, tag_to_xpath = await tarsier.page_to_text(page)

        print(tag_to_xpath)  # Mapping of tags to x_paths
        print(page_text)  # My Text representation of the page


if __name__ == '__main__':
    asyncio.run(main())

Keep in mind that Tarsier tags different types of elements differently to help your LLM identify what actions are performable on each element. Specifically:

[#ID]: text-insertable fields (e.g. textarea, input with textual type)
[@ID]: hyperlinks (<a> tags)
[$ID]: other interactable elements (e.g. button, select)
[ID]: plain text (if you pass tag_text_elements=True)

Local Development

Setup

We have provided a handy setup script to get you up and running with Tarsier development.

./script/setup.sh

If you modify any TypeScript files used by Tarsier, you'll need to execute the following command. This compiles the TypeScript into JavaScript, which can then be utilized in the Python package.

npm run build

Testing

We use pytest for testing. To run the tests, simply run:

poetry run pytest .

Linting

Prior to submitting a potential PR, please run the following to format your code:

./script/format.sh

Supported OCR Services

Google Cloud Vision
Amazon Textract (Coming Soon)
Microsoft Azure Computer Vision (Coming Soon)

Roadmap

Add documentation and examples
Clean up interfaces and add unit tests
Launch
Improve OCR text performance
Add options to customize tagging styling
Add support for other browsers drivers as necessary

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/tarsier}
}

DSamuelHodge/tarsier