LLM reasoning engines for unstructured dataset parsing

I recently ran into a mess of raw text, emails, and call notes while trying to map out nut-free restaurants in Chicago. The standard approach would be to spin up a database or write a custom script to tag everything, but the overhead feels unnecessary when you can just dump the chaos into the prompt and let the model structure the output. It turns the LLM into a transformation layer for unorganized data.

I saw a thread on Hacker News about using this exact method to track personal fatigue data, which raised a question: how are people actually managing the state of these "knowledge" projects? I’ve been treating my findings like actual code, keeping everything in a local repo so I can version control the outputs, but I worry about the scalability of just appending to massive text files. Is there a more robust way to manage these research logs without them turning into unreadable garbage?

I'm curious if anyone is using specific markdown schemas or structured templates to keep these logs sane, or if we're all just one giant context window away from a total loss of data integrity.

If you want to see the schema I used for the Chicago data:

https://github.com/joshuabremer/chicago-allergy-eats

And the original HN discussion that started this:

https://news.ycombinator.com/item?id=48605117