Mixing properties of the same entity in the same TextSimilaritySpace

[miraseus]

Hello 👋
I'm investigating a use case where my entity X has several properties including product and product_category. I also have a user entity which has a desired_product property. Is it possible to combine all of those properties into a single space?

It would be something like this:
product_space = TextSimilaritySpace(
text=[X.product, X.product_category, user.desired_product],
model=""
)

The reason for this is that the user can set their desired_product arbitrarily and they may write "Dairy products" rather than "milk", which is much closer to X.product_category.

[morkapronczay]

Hi @miraseus!

The general rule of thumb is 1 schema 1 space 1 field, but for your use-case I can currently suggest 3 different approaches, please find them below and feel free to ask any followup questions about each!

1. concat the product fields outside SL system into 1 SchemaField and put that into the space (drawback is the vector will be the vector of the aggregate text of the fields, but using contextualised embeddings from sbert this should not be that big of a problem). This results in a space with 1 field from each schema.

text_sim_space = TextSimilaritySpace([user.pref, product.product_and_category_concatenated])

2. Create multiple spaces and add them to the index. This works until we search for product (as in your usecase), but we cannot search for the user having 2 similar clauses in the query with different input (like search for a user with a text query in product.description and a different text query in product. product_category)- that is a bug we are working on at the moment. But your usecase should be working fine!

text_sim_space_1 = TextSimilaritySpace([user.pref, product. product])
text_sim_space_2 = TextSimilaritySpace([user.pref, product. product_category])

3. Create multiple schemas for product, so we can have 3 fields in the space! Think of schema in this case as a language describing a vector, not an entity - so an entity can have multiple vectors. The problem currently is when we query we can only have 1 instance in the find clause so this approach would assume 2 queries done on each schema instances (prod_1 and prod_2) and combed together. This development is planned already (find accepting a list of schema instances).

text_sim_space = TextSimilaritySpace([user.pref, prod_1.description, prod_2.category])

[miraseus]

Hi @morkapronczay. Thank you for your suggestions! I will test them.