scann

import numpy as np import scann # 1. Prepare your dataset (e.g., embeddings) # For demonstration, creating a random dataset of 1000 vectors, 128 dimensions each. dataset = np.random.rand(1000, 128).astype(np.float32) # 2. Build the ScaNN searcher # This example uses dot product distance for Maximum Inner Product Search (MIPS). # num_leaves: Number of leaves in the tree for partitioning. # num_leaves_to_search: Number of leaves to search at query time. # anisotropic_quantization_threshold: Parameter for Anisotropic Vector Quantization. searcher = scann.scann_ops_pybind.builder( dataset, num_neighbors=10, # Number of nearest neighbors to retrieve distance_measure="dot_product" ).tree( num_leaves=100, num_leaves_to_search=10 ).score_ah( dimensions_per_block=2, # Recommended for MIPS anisotropic_quantization_threshold=0.2 ).reorder( 100 # Rescore top 100 candidates to improve accuracy ).build() # 3. Define a query vector query = np.random.rand(128).astype(np.float32) # 4. Perform a search neighbors, distances = searcher.search(query) print(f"Query vector shape: {query.shape}") print(f"Dataset shape: {dataset.shape}") print(f"Found {len(neighbors)} neighbors: {neighbors}") print(f"Corresponding distances: {distances}")