estimateNutrition() (lib/ai/features/estimate-nutrition.ts) is now a
hybrid: for each ingredient, estimateGrams() (lib/ingredient-grams.ts)
converts its quantity+unit to grams where possible (weight units
exactly; volume units -- cup/tbsp/tsp/ml/l/etc -- via a 1ml~=1g water-
density approximation, documented as a real simplification but far
better than skipping them). Convertible ingredients get looked up in
USDA FoodData Central (lib/usda.ts, SR Legacy + Survey (FNDDS)
datasets -- the two suited to generic/raw ingredients, not Branded
packaged products or narrower Foundation) and summed. Whatever's left
(count-based units like "2 cloves", no USDA match, or USDA_API_KEY
unset entirely) goes through one AI call asking for just that
subset's total contribution, which sums directly with the USDA
totals -- no whole-recipe AI estimate to reconcile against.
Same exported signature and return shape as before
(NutritionEstimate / {perServing: {...}}), so every existing caller
(the nutrition route, meal-plan generation) gets more accurate
results automatically once USDA_API_KEY is set in Admin -> Settings,
with zero behavior change when it isn't.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Arnaud