You can “search logseq ‘foo’” and then ask MCP to figure out where in those I discussed about how MCP sux
first keyword search then play on LLM strengths processing text to find vaguely related information, not even synonyms
so in this case, you can make a finetuned LLM that takes in the recipe and turns the ingredients into concentrations (guess)
however, imho you have a different problem: unclear goal -> not defined evals + spinning on the idea forever, instead of like say 15 mins
so my suggestion: define goal where the subs would clearly help to surface desired recipes
define eval with a minimal recipe DB, with minimal I mean like 5, then 15 and finally 50 recipes
then spin up a version in chunks of 15 mins. Not months. meaning: “I need to turn these recipes to ingredients and amounts” 15 mins
“I have these recipes, how can I extract ingredients and amounts?”
“I need to devise a substitution algo, to calc 2d score for each recipe how well another recipe could substitute it” (or maybe better, by ingredients first, and in later workchunk by 2nd level ingredient -> recipes substitution scores)
which means: you can not code whole solution in 15 mins
but you can iterate twice with LLM, and test it
with minimal recipe DB it means you can give a ground truth manually, both 2d tables and your eval cases
which means you will not get stuck for months with the immense problem