A retrieval pipeline returns many semantically related passages, but the best evidence ranks too low for the final prompt. Which approach best fits?

1. Evaluate reranking after candidate retrieval.
2. Choose and evaluate an appropriate chunking strategy with metadata.
3. Preserve source provenance and validate answer support against retrieved evidence.
4. Align and version the embedding representation on both sides.