One of the most popular skills in Anagram Experiences is Recommend Products. This skill allows the AI to narrow down which products best fit a user's needs from a large catalog.
Our original approach used a straightforward flow: run a single query to retrieve the top 75 most relevant products, pass them to an ai model, and have the model select the best matches. This worked well for simple requests — a good recommendation was almost always somewhere in those top 75 results.
Where it fell short was scenarios that required multiple queries.
For example, imagine instructions like: "If Product A is recommended, also suggest a compatible accessory." Our old approach couldn't handle this because a single query can't return two different categories of products — especially when the second recommendation depends on what the first one turns out to be.
The new approach: intelligent, multi-step search
Our updated skill uses a reasoning model with access to tools that let it search the catalog multiple times and decide on its own when it has enough information to make a strong recommendation. The flow now looks something like this:
Search for Product A.
Evaluate the results. If nothing fits well, refine the query and search again.
Once a good match is found, run a second search focused on finding an accessory that pairs well with the selected Product A.
Evaluate those results the same way, re-searching if needed.
Once both products check out, recommend Product A and its matching accessory together.
The key difference is that the AI now controls the search loop. It can run as many queries as it needs, adapt based on what it finds, and stop only when it's confident in the results.
With this update, you can provide more complex recommendation instructions and trust that the skill will intelligently work through them to find the best products for your customers.