vdutts7 / gpt4V-scraper

AI agent that can SEE πŸ‘οΈ, control, navigate, & do stuff for you on your browser.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Logo Logo Logo

GPT-4V Web Agent

AI agent that can SEE πŸ‘οΈ, control, navigate, & do stuff for you on your browser.

Github


Table of Contents

    πŸ“ About
    πŸ’» How to build
    πŸ”§ Tools used
    πŸ‘€ Contact

πŸ“About

  • Automated web scraping tool for capturing full-page screenshots.
  • Utilizes Puppeteer with a stealth plugin to avoid detection by anti-bot mechanisms.
  • Designed for efficiency with customizable timeout settings.

πŸ’» How to build

Part 1: Screenshot + Scrape

  • Run npm i to install dependencies (Puppetteer libraries, see package.json for details).
  • Copy .env.template and rename this new file .env . Then add your OPENAI_API_KEYand save the file. Run source .env properly mount this into the environment.
  • Set up browser confguration to allow for websites that require login authentication (LinkedIn, Instagram, etc). Make sure you log in ahead of time, that way your browser agent can access content without problems. For paywalled sites, it is your choice but hey: https://removepaywall.com/<URL> . This is a GitHub project, not a moral essay, so decide for yourself and move on. I will say, however, that sites such as NYT, CNN, FOX, Guardian, etc., shouldn't be misrepresenting themselves as "news" when they're making you pay for truth. But I (and you probably if you're the type of person reading a Github project description) see nothing valuable in sites like those anyways that is worth scraping. For the best browser, install Chrome Canary (log into the website of choice before continuing this next step). Then reference it in snapshot.js as follows:
executablePath: '/Applications/Google\ Chrome\ Canary.app/Contents/MacOS/Google\ Chrome\ Canary',
userDataDir: '/Users/vdutts7/Library/Application\ Support/Google/Chrome\ Canary/Default',

Replace paths with respective locations specific to your device.

  • Run node snapshot.js "<URL>" . Insert any URL in place of <URL>.

Ex:

node snapshot.js "https://en.wikipedia.org/wiki/Devious_lick"

Wait for few seconds (adjust const timeout = 6000; if too slow), and snapshot.jpg will magically appear in root directory of project:

snapshot.jpg

Part 2: Image to Text Conversion

This next part is better than a lot of OCR software for common tasks- in my opinion.

Edit the following lines in python gpt4v_scraper.py, replacing with your own website URL and then a system prompt (command to the GPT-4V API) about what to scrape for. See my example:

# Running the function w/ a example website + prompt
result = web_capture_and_extract("https://vd7.io/experience", "Extract this dude's Experience info")
print(result)

You should see two things happen:

  • snapshot.jpg will be the screenshot of the scraped site
  • the console will display the text in the snapshot.jpg with additional context and answers to your prompt input
console output

compare it to the actual site (not bad tbh):

vd7 website for comparison Screenshot 2023-12-20 at 2 28 12β€―PM

Part 3: "AutoGPT" style prompting + Bing search

Chat with GPT-4V Web Agent in realtime to guide it across Bing search.

Run node web_agent.js :

πŸ”₯ GPT-4V Web Agent: Sup, ask me something!
You: 

For example:

chat with web agent

which matches the trending cryptos list on CoinMarketCap:

via CoinMarketCap

Here's the extended convo πŸ˜‚ :

(base) vdutts7@Vacbook-Vro gpt4V-scraper % node web_agent.js 
πŸ”₯ GPT-4V Web Agent: Sup, ask me something!
You: gimme the top 10 trending cryptos right now

GPT: {"url": "https://coinmarketcap.com/trending-cryptocurrencies/"}
Crawling 🐜  https://coinmarketcap.com/trending-cryptocurrencies/
GPT: The top 10 trending cryptocurrencies on CoinMarketCap right now are:

1. Bonk (BONK)
2. Loopring (LRC)
3. Internet Computer (ICP)
4. Cirus Foundation (CIRUS)
5. Coq Inc (COQ)
6. Shiba Inu (SHIB)
7. SATS (100SATs)
8. Worldcoin (WLD)
9. r a t s (Ordinals) (rats)
10. Bitcoin (BTC)

Please note that cryptocurrency markets are highly volatile, and the trending list can change frequently.
You: which ones looks smexy

GPT: It's important to clarify that the term "smexy," which is often a colloquial or slang term for "smart and sexy," doesn't really apply to cryptocurrencies. Cryptocurrencies are digital or virtual currencies and don't possess physical attributes that can be defined as attractive or not in the way that term suggests.

However, if you're asking about which cryptocurrencies on the list are potentially interesting or have seen significant attention, people often consider a few factors such as market performance, innovation, and the presence of a strong community or technological potential behind the cryptocurrency.

If we look at the trend data provided and the percentage changes in price over the past 7 days from the snapshot:

1. Bonk (BONK) has seen a remarkable 613.52% increase over 7 days.
2. Cirus Foundation (CIRUS) has grown by 107.72% over the same time frame.
3. Worldcoin (WLD) has appreciated by 105.97% over the last 7 days.

These significant increases might signal market excitement around these tokens, suggesting that some traders may find these trends to be metaphorically "smexy."

However, this should not be considered investment advice. Cryptocurrency investments can be highly speculative and carry a high level of risk. It's important to conduct thorough research or consult a financial advisor before making investment decisions.
You: 

πŸ”§Tools Used

GPT4Vision Puppeteer

πŸ‘€Contact

Email Twitter

About

AI agent that can SEE πŸ‘οΈ, control, navigate, & do stuff for you on your browser.


Languages

Language:JavaScript 81.2%Language:Python 18.8%