Definition
A bi-encoder framework that maps queries and document passages into a shared high-dimensional vector space to perform semantic retrieval via similarity search. Unlike sparse methods, it captures latent relationships by optimizing the inner product between query and passage embeddings, though it faces a trade-off between higher semantic accuracy and increased computational cost compared to lexical search.
Semantic vector similarity retrieval vs. keyword-based lexical retrieval.
"A magnetic field where concepts with similar meanings are physically pulled toward the same coordinates regardless of the specific vocabulary used."
- Bi-Encoder(Architectural Pattern)
- BM25(Sparse Baseline Alternative)
- Vector Database(Indexing Infrastructure)
- Maximum Inner Product Search (MIPS)(Mathematical Foundation)
Conceptual Overview
A bi-encoder framework that maps queries and document passages into a shared high-dimensional vector space to perform semantic retrieval via similarity search. Unlike sparse methods, it captures latent relationships by optimizing the inner product between query and passage embeddings, though it faces a trade-off between higher semantic accuracy and increased computational cost compared to lexical search.
Disambiguation
Semantic vector similarity retrieval vs. keyword-based lexical retrieval.
Visual Analog
A magnetic field where concepts with similar meanings are physically pulled toward the same coordinates regardless of the specific vocabulary used.