Conversational AI in 2 Ecommerce Scenarios

API
, ,
1

Hey everyone!

I’m a digital product designer that is currently getting a price comparison for smart home products developed (e.g. a site like Google Shopping, Honey).

While the first iteration (MVP) won’t include conversational AI, I plan to include it in my shopping experience a few months after launch.

The main use case I see at the moment for my product are:

  • A deep knowledge product expert and shopping assistant that can answer any questions about the products (e.g. details, reviews, specs etc) I have in my database. Some important tasks would include helping a shopper learn what specs to consider when purchasing a product, helping compare products, provide recommendations etc.

So far, I’ve created a Custom GPT for myself to test out some of the prompt engineering with a JSON file with 400 products and it’s worked fantastic!

The stretch/vision use case I see is:

  • Customer prompts that can do all of the above plus dynamically update the page, parse/analyze info from the page and perform actions like Siri on iOS
    e.g. “My budget is $500 for a smart door lock” and have CGPT update the Price filter with $500 max.
    e.g. “Set a price drop alert for this product (If the user is on a product details page) if it drops by $20” and have CGPT create a price drop alert for the customer.

A few questions I have are:

  1. What would a high-level development approach look like for each?
  2. How could I make this as financially viable as possible considering that I’m a solo-preneur bootstrapping this?
  3. Any other feedback, concerns, or considerations I should be thinking about?

Thanks so much in advance!

2

My thoughts:
Approach 1: group products into categories (price and product type) then run function call(s) for each SKU property in the user query, find all matching products, then append to the assistant response to the user. (slower, more expensive, but has a lot of functionality with gpt-4o).

Or approach 2: build a RAG pipeline by embedding all product properties (attach product id to each attribute embedding), embed the user query, then run a cosine similarity over your product database, then send the user either a ranked list of relevant products or the most similar product. (cheap, fast, a bit of dump pipe thus more work).