The data pipeline is part of the answer
Extraction, cleaning, structuring, and permission mapping happen before search. Errors in those steps later appear as weak or contradictory answers.
A robust pipeline shows when content was last processed, which version is valid, and why a document failed to index.
Retrieval needs a reviewable benchmark
Hybrid search, semantic ranking, and query rewriting are tools. Their effect should be measured against real questions and expected sources.
- Test questions from actual workflows
- Expected documents and acceptable alternatives
- Measures for retrieval quality, faithfulness, and latency
- Separate analysis by content type and user role
Permissions are part of relevance
A relevant result is unusable when the user is not allowed to see it. Identity and access rules must remain intact through the index and answer generation.