<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/6b34caf8-948e-4b31-96a8-07d0da1ebd86/a0aca5b0-eb58-4785-bd22-822625f9c18c/github-mark-white.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/6b34caf8-948e-4b31-96a8-07d0da1ebd86/a0aca5b0-eb58-4785-bd22-822625f9c18c/github-mark-white.png" width="40px" />
This is nutramap, the nutrition tracker of my dreams.
Nutramap has the perfect balance of convenience, flexibility, and accuracy. It uses natural language processing, so you can type what you had to eat yesterday just like you’re texting your mom about it, and nutramap breaks down your meal into ingredients, estimates the amounts of each, and maps each ingredient to the most comprehensive nutrition database there is,
What can’t be measured, can’t be managed - with a focus on seamless interface and convenience, Nutramap is the perfect solution to gain vital awareness over your own diet.
</aside>
Languages
This is nutramap, the nutrition tracker of my dreams.
Nutramap has the perfect balance of convenience, flexibility, and accuracy. It uses natural language processing, so you can type what you had to eat yesterday just like you’re texting your mom about it, and nutramap breaks down your meal into ingredients, estimates the amounts of each, and maps each ingredient to the most comprehensive nutrition database there is, the USDA’s. There are thousands upon thousands of foods in this database, every version of every raw ingredient you can imagine and many many branded foods as well, each with a full accompanying nutrition panel that has over 70 different nutrients.
For example, if I typed in yogurt, it could be any of these and more:
So parsing user input and mapping it as accurately as possible to an entry was no minor task, especially because you want to keep the appfluid and responsive as possible.
To do this, I combined dense and sparse indexing for each ingredient. FAISS for vector similarity search and Typesense for keyword-based search.
So each ingredient got put through each index, and the generated similarity scores were combined using using a Reciprocal Rank Fusion algorithm. This is similar to how most modern RAG pipelines work!
It takes into account both absolute similarity score and relative ranking, and - most importantly - it accounts for ties. Typesense returns equally ranked values in a somewhat random order, so before I adjusted the algorithm to take ties into account, I had this issue where very simple ingredients got comically incorrect results because so many entries got a perfect match with sparse indexing, to the point where the actual best match didn’t even make it into my list of top 50 results. For example, I would get “Pound cake made with butter” as the top result for just “butter”
Another way I kept calculation times low was by using multithreading whenever possible. Pretty much every loop and non-dependent set of statements in my backend is executed in parallel using python’s asyncio library.
One big problem was that FAISS is SLOW and expensive. I minimized both time and cost by saving the embeddings for each food in my Mongo database, caching the generated FAISS index in a .bin file,