Versions (relevant - OpenSearch/Dashboard/Server OS/Browser):
OpenSearch version : 2.11
Model used : sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 · Hugging Face
Describe the issue:
I have a database of common e-commerce products (clothes and accessories like tshirts, mugs, posters…). I’m trying to improve the existing search experience which use keyword search by adding semantic search.
I have selected and tested a model (see above) which works well when performing simple searchs (simple neural search on one specific field → category embedding or product name embedding).
However, when i try hybrid searchs i have got unexpected results.
Ex1 : Hybrid search combining neural search on product’s name and category name (with 50/50 of weight) : If i search for “Mug dad” i get results of products containing “dad” in their names but from any categories instead of having mugs first (i have multiple products that are supposed to match)
Since it’s an hybrid search with 50/50 weights i’m expecting that if i have products with dad in their name and from mug category, they should have higher scores and be on top of the ranking.
I have tried multiple configurations such as changing search pipeline settings (normalization and combination) but nothing gives me relevant results.
Anyone having experience with hybrid search for e-commerce database and willing to give advices ?
Configuration:
Relevant Logs or Screenshots: