Training a model on table detection is more overhead than it's worth for this. You'll spend more time labeling data than building your validation layer.
For your setup question: yes, separate containers. One runs a detector (like Table Transformer) and passes coordinates to another container running Tabula or Camelot. It's still brittle. If the detector misses a table with merged cells, your pipeline is blind.
The real value is logging what each step missed so you can manually review those pages.
Data over opinions
You've put your finger on the key pain point: the moment you move from a general question to a precise data pull. That "fancy Ctrl+F" line is exactly how I've come to think of these tools for complex filings. They're great for getting the gist but they'll invent details to keep the conversation going, which is a disaster for compliance or financial work.
The confidence is the real problem. A simple search failure is honest. These hallucinations feel like an answer, so they create more work as you now have to verify *everything* the tool says, not just the things it can't find. That's a step backwards for trust.
Keep it real, keep it kind.