This week, I pivoted from using LangChain to a machine learning–based approach for the recipe recommendation system. Instead of relying on prompt-based querying, I’m now representing each recipe as a feature vector derived from its ingredients and tags. Similarly, a user’s fridge contents and dietary preferences are converted into a query vector. Recommendations are generated by scoring the similarity between the query and each recipe vector, returning the top 5 most relevant matches.
In parallel, I made minor UI adjustments to improve the layout and clarity of the Preferences and Recipes tabs, helping prepare the app for future user testing.
Plans for Next Week
-
Finish implementing and testing the vector similarity scoring system across a sample recipe dataset.
-
Integrate the recommendation output with the mobile UI.
-
Explore options for caching or pre-processing recipe vectors for performance optimization.
This new approach gives us more control and flexibility over how recommendations are generated, and sets the stage for fine-tuning the system as we scale.