QuantizationSearchParams
@Serializable
How a search reads a quantized collection.
This is the half of quantization that decides recall, and leaving it unset is the mistake: a collection quantized to a quarter of its size answers from the approximation unless something asks for the originals, and a caller who never sets rescore has traded accuracy they did not choose to trade. Qdrant decides for itself when it is unset, and its choice depends on where the vectors are stored.
oversampling is the lever that makes rescoring worth it: preselect more candidates from the quantized index, then rank that larger set against the originals. At 2.0 with a limit of 100, Qdrant scores 200 approximately and returns the best 100 exactly.