yanoyan / cocktail_matcher Goto Github PK

This demo code helps you find your perfect cocktail recipes in this "not-so-dry" January. It leverages the Huggingface's dataset, cocktails_recipe_no_brand, which contains 6.6K cocktail recipes. The data fields title, glass, garnish, recipe, and ingredients are concatenated into combined_text and processed using OpenAI's text-embedding-ada-002 to build embeddings. The dataset and precomputed embeddings are stored in Neon's serverless PostgreSQL database.

The script uses function calling to extract desirable cocktail attributes from a given user input. The formated query is constructed by the extract (and enriched) cocktail attributes, formatted in the same way the Huggingface recipe texts are combined. The L2 distance is used for the similarity search to find the best-matched cocktail recipes. Note that function calling not only helps extract and structure the user input but also enrich it. For instance, if a user asks for "old fashion", the ingredients and recipe are auto-populated by OpenAI's API response.

Below is a simple interface of the Cocktail Matcher.

Topher said we are taking this little thing to a bar on Feburary 1st.