Ai Search _ lacking ML recommendation system knowledge using LLM models

[river salmon]

I’m a complete nooobie at ML, but I’m attempting to create a search integration with azure ai search and while it’s good at NLP and finding searched matches, it doesn’t incorporate the feedback loop for recommendations which I’d imagine would involve storing each users purchase history, best seller , etc. how do I begin to implement some sort of a feedback loop like a llm model and are there any tutorials for this type of thing?

[jade monolith]

Are you trying to have it act similarly to Amazon in which it will change it's recommendations based on a user's past search history?

[river salmon]

I am in a way, and also looking to create a classification model that ranks results with the highest likely hood of being purchased on the top. I know this is an advanced topic to cover over discord but a good starting point would be appreciated! Thanks

[jade monolith]

This wouldn't be something done purely by AI Search. I would recommend having your backend update a database of documents with how much they get queried and then add that field in your search queries. For per person recommendations, basically same thing but you'll make a list of search terms/etc which you can append to your search queries to help refine them.

[river salmon]

I believe I have a database of search term usage, thank you. Per user search terms are a great start as I already include an array of customer ids who’ve purchased the product to each product because Azure guidance on using multiple index’s is to not do that, (I.e. customer and product) like a relational database. I have Concerns over search performance of this type of usage, would you recommend another method for the entity linking in ai search?

[jade monolith]

You don't have to entity link the customer's search history with your data source at all tbh.

What you can do is have a field in your customer's data such as past_queries wherein you just append their latest search queries to it (after removing duplicates) and then append this field to your queries.

This could get long, so alternatively if you just want to get specific on items your customer was interested in, have a field in their data of past_items and it just be a list of item_id which you can then append to your search queries as an "or".

You'll have to experiment a bit since it will heavily depend on how you want to make this work for what you're building. You'll definitely also have to do a bit of heavy lifting on your own client.

Keep in mind that result that Scoring Profiles are likely the way to go, but you'll have to build one based on your customer: https://learn.microsoft.com/en-us/azure/search/index-add-scoring-profiles

[river salmon]

Looks good, thanks !