HN user
I had done the same for classifying and generating bookmarks of thousands of datasheets, along with a very naive yolo-based classificator (to detect pages made out of diagrams and pictures mostly).
Done with GLM-OCR, I had to watch text sloooowly crawl out of the llm and still have to live with hallucinations and the model not following the schema
"Press here to view the content", there's already plenty in the wild that grant access to notifications with deceptive buttons.
You "personally checked" the result (generated by an LLM, a huge black box with extensive knowledge of all fields) to the best of your knowledge. There is a mismatch between what the machine knows (and has done as the result of it) and what you think you know.
Implementing a fix implies knowledge of the inner workings that brought you to it. A fix made by a LLM does not give you that.
Wow. Incredible how this was not mentioned in the OP. I had done it with tcn triggers and adding "_archive" shadown tables manually with tcn (https://www.postgresql.org/docs/current/tcn.html), but doing it natively is gonna be, as per most postgresql implementations, wonderful.
The whole sqlite documentation is full of gold gems and other curious documents mostly to appease bureocrats and big companies. It doubles as a fun read other than being incredibly useful.
See, for example: "Defense about the dark arts" (https://sqlite.org/security.Html) and "Why in C?" saying "Because C is best."
Hello. Author here, I usually just lurk with no account. I wrote this post after fixing a machine that stopped working due to the SD degrading away. Last time it happened I had to redo everything from just carved data, so this time I tried to have fun and see how much i could salvage. There was a developer "backdoor" and weirdly aesthetic corruption that was too cool not to share