Semantic Search

Semantic search studies how to find relevant documents, products, code, and other data by meaning rather than relying only on matching keywords. Common approaches use dense embedding models and vector indexes, often combined with lexical search, structured metadata, or cross-encoder reranking; newer systems also use LLM-generated query expansions and relevance labels. Evaluations span benchmarks such as BEIR and code-retrieval datasets as well as deployed settings including e-commerce, clinical records, and job search, where latency, scale, and user relevance matter. A central challenge is balancing efficient retrieval with genuine meaning-level matches: models can miss equivalent ideas expressed differently, while richer reranking, privacy protection, and index construction add cost and complexity.