Wine Recommender

Describe your ideal wine and get relevant matches based on text similarity.

I’ve updated the search functionality to use more sophisticated semantic search methods leveraging (sentence) transformer models, which I find yields excellent recommendations that closely match search queries. So far, the models I’ve tested are mpnet-base-v2, msmarco-MiniLM-L-6-v3, msmarco-distilbert-base-v4. They all seem to perform really well.

Users can pick countries, minimum points given by Winemag and the number of matches to return.

Besides wine names and descriptions, users may to look up any recommended bottle on wine-searcher.com by clicking the links provided in the output.

I’m quite happy with how well results match the queries if tested, but for my own curiosity’s sake, I might try and improve results by applying generative pseudo labeling for domain adaptation.

Check it out here!

Tools and technologies
  • Python
  • Transformers
  • Semantic search
  • Shiny
  • Docker