Retail API, anyway to fine-tune/train semantic search ?

[garyz]

I am evaluating retail search. With regard to semantic search, I saw some out-of-the-box impressive features, for example. the famous search phrase:  long sleeve  short baby  shower dress.

But each application is different, there are some specifics. Application developers are happy to provide sample data to train (or fine-tune) the model, anyway to do that ?

An hypothetical example, products like below:

    2022 Chevy Camaro, with mileage of 8,000, and 20 MPG
    2015 Chevy Cruze,  with mileage of 40,000, and 32 MPG

The search terms such "less than 5 year old cars", "under 10,000 miles", or "30 MPG or higher" do not work, I tried.

Anyway to do 'few-shots" training?   we are happy to provide data, though it should be only for our application.

[Joevanie]

You could collect examples of search queries like "2022 Chevy Camaro less than 5 year old cars", "2015 Chevy Cruze under 10,000 miles", and "30 MPG or higher cars". Then, pre-train a large language model (LLM) on a large corpus of text data. This will give the LLM a general understanding of language. Then, fine-tune the LLM on the few-shot data. This will teach the LLM to associate specific search queries with specific products.

[garyz]

Thanks for replying, Joevanie.   I am testing Google retail search as a product,  but couldn't find any interface to train LLM part of the retail search.   Anyway we can integrate with other LLM services which we can train?   Or any train/fine-tune interface for retail search I am not aware of?  I'd appreciate if you can point me to the direction.

I am looking into Retail search as a possible one-stop solution,  so we don't have to build several components for search, personalization and recommendations, based on tracked success conversion rates.   All of these are coming out of the box.