Here are some details since you asked! Its actually using perplexity sonar under the hood and running a search across a few databases (the same ones myfitnesspal and others use) its extremely expensive and costs me almost 1 cent every time someone logs food 😅 Unfortunately it is a situation where the quality of the input will define the quality of the output So if you type in something like pizza, it's going to have a much harder time understanding and it's going to try to make a best guess. Honestly similar to just typing pizza in myfitnesspal and selecting the first option If you type in something like "two slices of pizza from Pizza Hut," then the accuracy goes up tremendously. And then obviously if you type in something as detailed as you would in My Fitness Pal, like "13.4 oz of ground beef from Whole Foods with 3 flour tortillas also from Whole Foods" It is extremely accurate I'm using My Fitness Pal as my benchmark because it's one of the best apps in terms of databases. I actually have an internal benchmark that I've run internally of about a hundred different items in different categories and I compare the results of Amy against the best option in MyFitnessPal and Amy actually performs better overall in the benchmark (probably because im paying 5-10x for fresh data than they are) I'm working on cleaning up the benchmark and publishing it so other people can audit and help me improve Amy. Happy to share it with you once it's out next week :) EDIT: ill just publish the benchmark now, but will try to clean it up a bit to make it even better https://amyfoodjournal.com/accuracy
Hey Mike appreciate you taking the time. Would you be able to share an example of incorrect data? Also noted on the trial length, I'm starting to get the costs a bit more down so maybe I can look to try to extend it now^
Pizza can never be just 200 ;-) For the AI are you using openAI for example or manage to run a open source library llama for example?