Providing product recommendations seems like it should be simple, right? You should be able to just run a search through your catalog and boom; here's what you're looking for. Turns out there's a lot more to it than that, and shoppers often see irrelevant results.
How most product recommendation systems work
Most chatbot product recommendation algorithms are pretty simplistic: they either just use a basic product keyword search (similar to what you see on most e-commerce sites), or if they've put in a little more effort they have a RAG (retrieval-augmented generation) based setup that allows for product searches based on semantics rather than pure keywords.
Okay so you found some products that maybe match, now what?
The problem with both of these approaches is that while they can retrieve potentially relevant products, they don't employ an additional filtering mechanism to figure out which of those products actually match what the user was looking for. What we've done is implement an additional reasoning-based filter to select which products from the initial query are actually relevant.
As an example: perhaps you're shopping for plants and we have the context that you want something with smaller leaves. That might not be something you can capture in the initial query, but an LLM (large language model) can probably make a determination after the initial selection process of which results have smaller leaves.
Weaving in additional context
Giving systems as much data as you can about preferences and past purchases is essential to providing effective recommendations. If you already purchased a pair of running shorts one week ago and now you're looking for something else, it doesn't make sense to recommend those same shorts to you; you already know about them.
Being careful about hallucinations
If there's one thing we all know about AI by now, it's that it's prone to inventing information if it can't find anything relevant. Addressing this problem head-on by actively managing product search outside of the LLM layer is essential to making sure that you're only presenting data to users that is actually backed by an external source.
Want to chat with us about boosting your product recommendations? Reach out.