Scott Horvath's Weblog

6-min read

A 'sophisticated' Facebook Marketplace alert system using GitHub Actions and ChatGPT's API

A scene from The Castle

If The Castle (1997) was filmed today, would Steve Kerrigan instead be endlessly scrolling through Facebook Marketplace and forever sending screenshots of jousting sticks to the Kerrigan family WhatsApp group?

The Facebook Marketplace app includes a feature that notifies you when a listing matches a specified keyword.

Unfortunately this feature is poorly executed, appearing more intent on driving app engagement with false positive alerts than truly helping users find what they’re searching for.

Defining the problem

The Facebook Marketplace ‘new listing’ notification feature:

The solution

Drawing inspiration from Simon Willison’s brilliant post on (and coining of) Git scraping, I built an open source alert system running on GitHub Actions and applied ‘innovative’ filtering using ChatGPT’s API.

A process flow illustrates a complete solution operating on GitHub Actions. A cron job activates a Python script every few hours to scrape recent listings. These listings are processed through basic logic and then analysed by the ChatGPT API as a final step to discern additional attributes that contribute to the alert criteria. If a listing matches, an email notification is triggered.

Example: Hali Rugs

This configuration will monitor and alert for Hali rug listings that are 2.8m by 2.3m with a 20% tolerance in dimensions.

Basic filtering (shown in the first yellow diamond) is applied to ensure that either the listing title or description contains the word 'hali'.

This is required because Facebook Marketplace will sometimes insert ‘relevant’ listings that don’t necessarily match the specific keyword, and so these can be filtered out using basic string matching.

Side note: a simple string filter is also a way to save on usage costs of the ChatGPT API in the next step.

def heading_details_keyword(details_collected_text, title_collected_text):
    # Check if the lowercase text contains 'hali'
    essential_word = 'hali'
    if essential_word in text_lower or essential_word in title_lower:
        return True   
    else:
        return False

Then (shown in the second yellow diamond), the ChatGPT API is used to transform the listing title and description into a more structured format that can be used by Python.

The prompt sent to the ChatGPT API is shown below. It also defines how the API must return a response.

prompt = f"""
Based on the following description and title for an item listed on Facebook Marketplace, determine if the item is a rug or floor runner. 
Respond strictly in the format 'yes|d1|d2' if it is a rug or floor runner, with 'd1' and 'd2' as the dimensions in meters. 
If the item is not a rug or floor runner, respond with 'no'. If dimensions cannot be determined, use 'na' for 'd1' and 'd2'.

Description: {description}
|||
Title: {heading}

Note: Your response should strictly follow the 'yes|d1|d2' or 'no' format without additional explanations.
"""

The dimensions can then be tested against the target dimensions.

# Calculate the minimum and maximum dimensions with tolerance for both width and length
min_width = desired_width * (1 - tolerance)
max_width = desired_width * (1 + tolerance)
min_length = desired_length * (1 - tolerance)
max_length = desired_length * (1 + tolerance)

# Check if either combination of dimensions are within the desired range
if ((min_width <= dimension1 <= max_width and min_length <= dimension2 <= max_length) or
   (min_width <= dimension2 <= max_width and min_length <= dimension1 <= max_length)):
    return True
else:
    return False

If the listing meets the target dimensions, an email notification containing the URL to the Facebook listing is triggered.

# Email body - HTML version
html = f"""\
<html>
  <body>
    <p>Hi,<br>
       Check out this heli rug item with dimensions we want: <a href="https://www.facebook.com/marketplace/item/{item_id}">Item Link</a>
    </p>
  </body>
</html>
"""

Example: Grange Furniture

This configuration will monitor and alert for any Grange furniture listings.

Unfortunately, there is a furniture store located on Grange Rd, Cheltenham and some other furniture is listed as Grange style. This would otherwise result in false-positive listings making it through to the second-stage filter and racking up unnecessary ChatGPT API usage costs.

The preliminary filter below ensures that API costs are kept to a minimum.

def details_are_exclude(details_collected_text):
    # Convert the collected text to lowercase for case-insensitive comparison
    text_lower = details_collected_text.lower()
    keywords = ['grange rd', 'grange road', 'near grange', 'grange style', 'grange-style', 'grange view']
    # Check if the lowercase text contains
    return any(keyword in text_lower for keyword in keywords)

Next, given the term 'grange' is popular amongst wine and Holden Special Vehicle (HSV) enthusiasts, the ChatGPT API is used to indicate whether the listing is a piece of furniture.

prompt = f"""
Consider the following description and title for an item listed on Facebook Marketplace. 
Your task is to determine if the content suggests that the item is NOT a piece of furniture. 
The item description is provided first, followed by the title, separated by '|||' for clarity.
Description: {description}
|||
Title: {heading}
Based on the description and title, is the item a piece of furniture? Please respond with 'yes' if it is a piece of furniture, and 'no' otherwise.
"""

End

Feel free to fork and modify either repository (linked in the examples above) to set up your own Facebook Marketplace alerts!

← Prev Next →