Query rewriting is the stage I would trace first in a web-search setup, because it is where the constraints quietly disappear. The user asks about a specific version or a specific year, the rewriter produces a cleaner general query, results come back genuinely relevant to the topic and wrong for the question - and every span downstream is green because each stage did its job on the input it was handed. Logging original question next to effective query makes that a one-line diagnosis instead of an afternoon. Web search adds a second thing worth tracing that internal corpora do not have: provenance of the fetched page. Retrieved web content is untrusted input, so it is worth recording which domain each selected passage came from and whether anything instruction-shaped survived extraction. Deduplication is also load-bearing here in a way people underestimate - the same syndicated article on six sites reads as six independent confirmations to the model, and consensus is not the same as corroboration.
Query rewriting is the stage I would trace first in a web-search setup, because it is where the constraints quietly disappear. The user asks about a specific version or a specific year, the rewriter produces a cleaner general query, results come back genuinely relevant to the topic and wrong for the question - and every span downstream is green because each stage did its job on the input it was handed. Logging original question next to effective query makes that a one-line diagnosis instead of an afternoon. Web search adds a second thing worth tracing that internal corpora do not have: provenance of the fetched page. Retrieved web content is untrusted input, so it is worth recording which domain each selected passage came from and whether anything instruction-shaped survived extraction. Deduplication is also load-bearing here in a way people underestimate - the same syndicated article on six sites reads as six independent confirmations to the model, and consensus is not the same as corroboration.