Definition
The mathematical process of rescaling embedding vectors to a unit length of 1, ensuring that retrieval is based solely on angular direction rather than magnitude. While it optimizes search performance by enabling faster Dot Product operations, it discards vector magnitude which occasionally carries semantic density or confidence signals.
Geometric scaling of embeddings for similarity search, not relational database normalization (3NF).
"Shrinking or stretching arrows of various lengths until they all touch the surface of a single sphere with a radius of exactly one."
- Cosine Similarity(Target Metric)
- L2 Norm(Mathematical Mechanism)
- Dot Product(Search Algorithm Optimization)
- Euclidean Distance(Alternative Metric)
Conceptual Overview
The mathematical process of rescaling embedding vectors to a unit length of 1, ensuring that retrieval is based solely on angular direction rather than magnitude. While it optimizes search performance by enabling faster Dot Product operations, it discards vector magnitude which occasionally carries semantic density or confidence signals.
Disambiguation
Geometric scaling of embeddings for similarity search, not relational database normalization (3NF).
Visual Analog
Shrinking or stretching arrows of various lengths until they all touch the surface of a single sphere with a radius of exactly one.