Definition
The process of transforming high-dimensional vector embeddings into a lower-dimensional representation to reduce computational overhead, storage costs, and search latency in vector databases while attempting to preserve the semantic structure of the data.
In RAG, this is not about compressing file sizes like a .zip; it is about reducing the number of numerical features (the vector length) used to represent a text chunk.
"Flattening a 3D topographic model into a 2D contour map: you lose the physical depth, but the essential boundaries and shapes remain recognizable."
- Vector Embeddings(Prerequisite)
- Cosine Similarity(Component)
- Approximate Nearest Neighbor (ANN)(Component)
- Information Loss(Trade-off)
Conceptual Overview
The process of transforming high-dimensional vector embeddings into a lower-dimensional representation to reduce computational overhead, storage costs, and search latency in vector databases while attempting to preserve the semantic structure of the data.
Disambiguation
In RAG, this is not about compressing file sizes like a .zip; it is about reducing the number of numerical features (the vector length) used to represent a text chunk.
Visual Analog
Flattening a 3D topographic model into a 2D contour map: you lose the physical depth, but the essential boundaries and shapes remain recognizable.