For ANY "our RAG is broken" question, your clarifying volley should hit these four buckets:

• The fix: detect tables during parsing (Document AI Layout Parser flags them), then either:
• Numbered sections and sub-sections
• Footnotes
• "Healthcare + wrong dosage" ← this is a safety-critical domain.
• specific details in the prompt
• Incremental indexing
• prompt caching
• When a customer says "this got expensive at scale" the FDE diagnosis tree is always:
• Parse the prompt for signal — when the question contains "tables," "drug dosages," "$5M," "2023 vendor," those are specific hooks. Address each one explicitly. You're missing one or two per answer; closing that gap is worth 20% on quality.
• Lead with diagnosis before tools — "naive RAG fails here because X" is a more impressive opener than "I'd use tool A and tool B." The diagnosis sentence is where the FDE judgment shows.
• Reframe the question when it has a wrong premise — Q3 is the classic example. "What do we cut" assumes cutting. The pro move is "let's measure differently." Watch for this pattern; it comes up constantly in customer conversations.
• Vocabulary precision — small but compounds. "Retrieval method" not "mode of chunk retrieval." "Parent-document retrieval" not "parent document strategy." Interviewers register the precise terms.