SBERT models https://www.sbert.net/ are trained to build a similarity embedding between a query and a passage.
The problem is that queries and passages are very different in length, form and semantic.
A solution with HyDE ?
Why not instead match a generated query passage from a model like OpenAI #GPT to passages in a vector database like Weaviate: compare the generated query passage embedding to (near) all database embeddings?
Another module for Weaviate in preparation?