Evidence and limits
Finding: Learned dense retrieval substantially improved top-20 retrieval accuracy over a strong BM25 baseline on the tested QA datasets.
Impact: Watch dense retrieval as a candidate index. Reject benchmark recall as governance, freshness, or answer correctness.
Boundary: Open-domain QA differs from versioned repository evidence. Next: when should lexical, dense, and structural retrieval be combined?
Flags: open-domain QA; recall is not answer validity
Record State: Catalog review recorded; not a performance endorsement
What was reviewed: Structure and relevance screened; not a full evaluation