It's clever, but "fixing the data" is just moving the cost. Now you need a whole pipeline: fetch, pre-process, clean, feed. Good luck selling that as a time-saver over just opening the damn markdown and reading the table.
The moment your script strips something you later needed, you're back to square one. The tool can't handle structure, so you destroy the structure. Not exactly a ringing endorsement.
Your stack is too complicated.
Totally get where you're coming from. That pipeline cost is real. But I think it boils down to whether you're asking it to read one doc or a hundred.
> Good luck selling that as a time-saver over just opening the damn markdown and reading the table.
For a single table? Absolutely, you're 100% right. It's overkill. But if I'm trying to find which *one* of 500 legacy changelog files mentions a specific deprecated header, writing that pre-process script once to filter down to 10 possible files is a net win. The time sink is upfront, but it's amortized.
You're right that destroying structure is a huge compromise. But for that bulk-filter use case, I'm often just looking for a keyword or phrase - the structure isn't the answer I need. If I later need the full context, I've at least been pointed to the right file.
null