Semantic Search

A dense semantic retriever encodes a query and searchable items into compatible vectors, then uses a specified score to select candidates. This can retrieve related expressions without identical wording, although success depends on the trained representation and query. Lexical systems emphasize term-based evidence; describing every keyword system as literal exact matching overlooks possible stemming, synonyms, and other processing. The two approaches therefore need an explicit comparison rather than a claim that one always understands meaning and the other never does. Dense Passage Retrieval is a concrete dual-encoder method for open-domain question answering. It trains question and passage representations so that a relevant passage scores above sampled negative passages, using vector dot products. At inference, its passage vectors are computed and indexed offline; a new question is encoded at query time. The distinction lets a searchable corpus be prepared without knowing every future question. An approximate index such as HNSW can reduce search work by navigating a hierarchy of proximity graphs. Its speed and recall depend on configuration and data, and approximation does not guarantee exact neighbors. A separate reranking stage can jointly process the query with each selected candidate. This cross-encoder design supplies pair-specific evidence at a different computational cost from comparing reusable vectors. It is not guaranteed to improve every dataset. Hybrid retrieval combines lexical and semantic signals, with the combination rule and evaluation making the comparison meaningful. In a RAG application, retrieved passages become context for an answer generator. Selection, source relevance, and answer correctness remain different outcomes; a vector score alone certifies none of them.