Quantized Clustering configuration
QcBuilder
Parameter | Type | Default value | Description |
qc.builder.train_sample_count | uint32 | 0 | The amount of training data. If this parameter is set to 0, all data is used. |
qc.builder.thread_count | uint32 | 0 | The number of threads enabled during index building. If this parameter is set to 0, the number of CPU cores is used. |
qc.builder.centroid_count | string | Optional | The centroid parameter. Hierarchical clustering is supported. Separate levels with an asterisk (*). Example of one-level clustering: 1000 Example of two-level clustering: 100*100 If you use two levels of centroids, better results are generally achieved when the first level has more centroids than the second level. Empirically, the first level has 10 times as many centroids as the second level. If this parameter is not configured, the system automatically derives an appropriate number of centroids. We recommend that you let the system derive the number automatically. |
qc.builder.quantizer_class | string | - | The quantizer to use. By default, no quantizer is used. Valid values: Int8QuantizerConverter, HalfFloatConverter, and DoubleBitConverter. Configuring a quantizer generally improves performance and reduces the index size, but may reduce recall to a degree that depends on the scenario. |
qc.builder.quantize_by_centroid | bool | False | Specifies whether to perform quantization by centroid when proxima.qc.builder.quantizer_class is used. This parameter is supported only when proxima.qc.builder.quantizer_class is set to Int8QuantizerConverter. |
QcSearcher
Parameter | Type | Default value | Description |
qc.searcher.scan_ratio | float | 0.01 | Used to calculate the value of max_scan_num, which equals the total number of documents multiplied by scan_ratio. |
qc.searcher.brute_force_threshold | int | 1000 | If the total number of documents is less than this value, linear retrieval is used. |