Why do exact terms remain valuable?

Identifiers, policy numbers, error codes, and specialist names can determine the right result. Semantic similarity may retrieve a related topic but miss an exact value. Sparse retrieval keeps these lexical signals explicit.

How do sparse indexes retrieve documents?

A document is represented by weighted terms. An inverted index maps terms to documents, enabling efficient candidate lookup. BM25 weighs term matches; learned sparse models can expand and weight features. Tokenization matters, especially for Chinese without word separating spaces. Multimodal retrieval and analytics shows retrieval in a larger workflow.

Dense retrieval bridges paraphrases, while sparse search excels at exact and rare terms. Hybrid search merges candidates and may rerank them. Choose a combination by measuring real queries, latency, and evidence recall.

FAQ

Is BM25 all of sparse retrieval?

No. Learned sparse representations are another approach.

Why does Chinese need tokenization care?

Segmentation changes the indexed terms and matches.

Can sparse retrieval support RAG?

Yes, especially for codes and proper names.