ScalarQuantizer<T>
Models & Types · AiDotNet.RetrievalAugmentedGeneration.VectorSearch.Quantization
Per-dimension scalar quantizer that maps each floating-point component to a single unsigned byte (uint8) using learned min/max ranges. Provides roughly 4x memory reduction versus 32-bit floats (8x versus 64-bit doubles).
Implements: IVectorQuantizer<T>
For Beginners
Imagine each dimension has values between, say, -3 and +5. We slice that range into 256 equal buckets and store just the bucket number (0-255, one byte) instead of the full number. To decode we return the middle of the bucket. Close, but far smaller.
How It Works
Each dimension is linearly mapped from its observed [min, max] range onto the
integer range [0, 255]. Reconstruction reverses the mapping using the bin center.
This is the same idea used by FAISS ScalarQuantizer (SQ8) and pgvector’s
halfvec/quantization paths.
Constructors
| Constructor | Summary |
|---|---|
ScalarQuantizer() |
Initializes a new instance of the ScalarQuantizer class. |
Properties
| Property | Summary |
|---|---|
bool IsTrained { get; } |
|
int CodeLength { get; } |
|
int Dimension { get; } |
Methods
| Method | Summary |
|---|---|
Vector<T> Decode(byte[] code) |
|
byte[] Encode(Vector<T> vector) |
|
void Train(IEnumerable<Vector<T>> vectors) |