/**
 * INT8 Quantization + batched cosine similarity search
 * Extracted from memory-initializer.ts (ARCH-4)
 *
 * The HNSW ANN index itself now lives inside @monoes/memory's
 * SqlBackend.search() (size-gated, built from the real memory_embeddings
 * table, persisted to disk next to the SQLite file) — see
 * memory-bridge.ts's bridgeGetHNSWStatus()/bridgeForceBuildHNSW(). This file
 * used to also maintain a second, separate pure-JS HNSW index reaching
 * directly into a `.swarm/memory.db` that nothing writes to post-rename;
 * that layer queried the wrong database with the wrong schema and was
 * removed rather than fixed in place, since the real search path already
 * covers the same need at the correct layer.
 *
 * @module v1/cli/hnsw-operations
 */
/**
 * Quantize a Float32 embedding to Int8 (4x memory reduction)
 * Uses symmetric quantization with scale factor stored per-vector
 *
 * @param embedding - Float32 embedding array
 * @returns Quantized Int8 array with scale factor
 */
export declare function quantizeInt8(embedding: number[] | Float32Array): {
    quantized: Int8Array;
    scale: number;
    zeroPoint: number;
};
/**
 * Dequantize Int8 back to Float32
 *
 * @param quantized - Int8 quantized array
 * @param scale - Scale factor from quantization
 * @param zeroPoint - Zero point (usually 0 for symmetric)
 * @returns Float32Array
 */
export declare function dequantizeInt8(quantized: Int8Array, scale: number, zeroPoint?: number): Float32Array;
/**
 * Compute cosine similarity between quantized vectors
 * Faster than dequantizing first
 */
export declare function quantizedCosineSim(a: Int8Array, _aScale: number, b: Int8Array, _bScale: number): number;
/**
 * Get quantization statistics for an embedding
 */
export declare function getQuantizationStats(embedding: number[] | Float32Array): {
    originalBytes: number;
    quantizedBytes: number;
    compressionRatio: number;
};
/**
 * Batch cosine similarity - compute query against multiple vectors
 * Optimized for V8 JIT with typed arrays
 * ~50μs per 1000 vectors (384-dim)
 */
export declare function batchCosineSim(query: Float32Array | number[], vectors: (Float32Array | number[])[]): Float32Array;
/**
 * Softmax normalization for attention scores
 * Numerically stable implementation
 */
export declare function softmaxAttention(scores: Float32Array, temperature?: number): Float32Array;
/**
 * Top-K selection with partial sort (O(n + k log k))
 * More efficient than full sort for small k
 */
export declare function topKIndices(scores: Float32Array, k: number): number[];
/**
 * Batched cosine similarity search with softmax-weighted scoring.
 * Combines batch similarity, softmax, and top-k in one pass.
 * Returns indices and normalized weights (not a Flash Attention kernel —
 * see the module-level note above).
 */
export declare function flashAttentionSearch(query: Float32Array | number[], vectors: (Float32Array | number[])[], options?: {
    k?: number;
    temperature?: number;
    threshold?: number;
}): {
    indices: number[];
    scores: Float32Array;
    weights: Float32Array;
};
//# sourceMappingURL=hnsw-operations.d.ts.map