QuantizedFlatIndex<T>
Models & Types · AiDotNet.RetrievalAugmentedGeneration.VectorSearch.Indexes
A flat (brute-force) index that stores compressed quantized codes instead of full
vectors, trading a small amount of accuracy for large memory savings and faster
scanning. This mirrors FAISS IndexPQ/IndexScalarQuantizer and Qdrant’s
on-disk quantized collections.
Implements: IVectorIndex<T>
For Beginners
This index keeps a tiny “summary” of each vector instead of the whole thing, so it uses far less memory. You still call Add and Search the same way; just remember to call Train once (on a representative sample) so it can learn how to summarize. Until then it behaves like an ordinary exact index.
How It Works
Vectors added before the underlying quantizer is trained are held in raw form and
searched with an exact fallback (equivalent to a FlatIndex). Once
Train is called, those vectors are encoded to compact codes, the raw
copies are released, and subsequent additions are encoded immediately.
Ranking after training uses the quantizer to score candidates: for a
ProductQuantizer paired with a distance metric it uses fast
Asymmetric Distance Computation (ADC); otherwise it reconstructs each candidate with
Decode and applies the supplied metric.
Constructors
| Constructor | Summary |
|---|---|
QuantizedFlatIndex(ISimilarityMetric<T> metric, IVectorQuantizer<T> quantizer) |
Initializes a new instance of the QuantizedFlatIndex class. |
Properties
| Property | Summary |
|---|---|
IVectorQuantizer<T> Quantizer { get; } |
Gets the quantizer used to compress vectors. |
bool IsTrained { get; } |
Gets a value indicating whether the underlying quantizer has been trained. |
int Count { get; } |
Methods
| Method | Summary |
|---|---|
List<ValueTuple<string, T>> Search(Vector<T> query, int k) |
|
bool Remove(string id) |
|
void Add(string id, Vector<T> vector) |
|
void AddBatch(Dictionary<string, Vector<T>> vectors) |
|
void Clear() |
|
void Train(IEnumerable<Vector<T>> vectors) |
Trains the underlying quantizer, then encodes and compacts any vectors that were added beforehand. |