Definition
The retrieval of data based on the mathematical proximity between high-dimensional numerical representations (embeddings), allowing RAG pipelines to find contextually relevant information by semantic meaning rather than literal keyword matches. It typically involves trade-offs between search latency, memory usage, and recall accuracy depending on the indexing algorithm used.
Semantic retrieval based on meaning, not lexical retrieval based on word frequency (BM25).
"A 3D star map where stars (data points) located physically closer together share similar characteristics, regardless of their names."
- Embeddings(Prerequisite)
- Cosine Similarity(Component)
- HNSW (Hierarchical Navigable Small Worlds)(Performance Optimization)
- K-Nearest Neighbors (k-NN)(Algorithmic Foundation)
Conceptual Overview
The retrieval of data based on the mathematical proximity between high-dimensional numerical representations (embeddings), allowing RAG pipelines to find contextually relevant information by semantic meaning rather than literal keyword matches. It typically involves trade-offs between search latency, memory usage, and recall accuracy depending on the indexing algorithm used.
Disambiguation
Semantic retrieval based on meaning, not lexical retrieval based on word frequency (BM25).
Visual Analog
A 3D star map where stars (data points) located physically closer together share similar characteristics, regardless of their names.