Sparse Retrieval

You would think that sparse retrieval is objectively worse than dense retrieval. But there are times where exact term matching matters a lot.

Dense models often smooth away rare, exact terms — because their goal is to encode meaning, not literal token identity.

Example:

Query:

“Error 0x80070005 Windows 11 update fix”

Dense embeddings will focus on the general meaning of “Windows update error fix”,
but the crucial part is “0x80070005” — a literal error code.