Hybrid RAG combines dense vector search with keyword search.
Dense catches paraphrase and conceptual matches.
Sparse catches exact names, codes and rare terminology.
Results are fused, most often with Reciprocal Rank Fusion.
It is close to the default choice for production systems now.
Hybrid RAG retrieves with both vector search and keyword search, then merges.
It is close to the default for production systems now.
The two methods fail in opposite ways, which is what makes the combination work.
Dense search understands paraphrase and misses exact identifiers.
Keyword search nails identifiers and misses synonyms entirely.
Real questions frequently contain both a description and a specific term.
How do I fix error XR-4471 needs the concept and the exact code.
Dense alone returns pages about similar errors, and keyword alone misses the explanation.
Results are merged with reciprocal rank fusion, which needs only the rankings.
That avoids comparing a BM25 score with a cosine score, which are on unrelated scales.
Weighting can favour one retriever when your corpus clearly leans one way.
The cost is two indexes to build, update and keep consistent with each other.
A document deleted from one and left in the other produces baffling behaviour.
Many vector databases now offer both plus fusion in a single query.
Add a reranker on top and you have the standard production shape.
One question on this card. Every option explains itself once you pick, right or wrong.
19 concepts in this module, then a graded test with every answer explained.