/**
 * HNSW Parameter Presets
 *
 * These presets are workload-informed starting points. `expectedRecall` is a
 * heuristic field from historical benchmark runs, not a guarantee; callers
 * should validate recall against exact brute-force ground truth on their data.
 *
 * Key parameters:
 * - M: Maximum connections per node (higher = better recall, more memory)
 * - efConstruction: Beam width during index building (higher = better quality, slower build)
 * - efSearch: Beam width during search (higher = better recall, slower search)
 */
export interface HNSWPreset {
    name: string;
    description: string;
    M: number;
    efConstruction: number;
    efSearch: number;
    expectedRecall: number;
    targetDimensions: string;
    targetDatasetSize: string;
}
export interface BrowserPresetOptions {
    /** Expected number of vectors. Used by getRecommendedPreset before browser caps are applied. */
    datasetSize?: number;
    /** User agent override. Defaults to navigator.userAgent in browsers. */
    userAgent?: string;
    /** Hardware concurrency override. Defaults to navigator.hardwareConcurrency in browsers. */
    hardwareConcurrency?: number;
    /** Force low-memory/mobile-style caps. Auto-detected from userAgent and hardwareConcurrency when omitted. */
    memoryConstrained?: boolean;
    /** Force Safari-specific search cap. Auto-detected from userAgent when omitted. */
    safari?: boolean;
    /** Force mobile-specific caps. Auto-detected from userAgent when omitted. */
    mobile?: boolean;
}
/**
 * Preset for low-dimensional vectors (128D or less)
 * Suitable for: Image features, word2vec, GloVe embeddings
 */
export declare const PRESET_LOW_DIM: Readonly<HNSWPreset>;
/**
 * Preset for medium-dimensional vectors (256-512D)
 * Suitable for: Sentence embeddings, smaller transformer outputs
 */
export declare const PRESET_MEDIUM_DIM: Readonly<HNSWPreset>;
/**
 * Preset for high-dimensional vectors (768D+)
 * Suitable for: BERT, GPT embeddings, Cohere, OpenAI embeddings
 * This is the recommended preset for RAG applications
 *
 * A seeded synthetic 768D sweep informed this heuristic, but recall remains
 * workload-specific. Validate the selected parameters against exact ground
 * truth for the target data before treating the preset as sufficient.
 */
export declare const PRESET_HIGH_DIM: Readonly<HNSWPreset>;
/**
 * Preset for very high-dimensional vectors (1536D+)
 * Suitable for: OpenAI text-embedding-ada-002, text-embedding-3-large
 *
 * Scaled from PRESET_HIGH_DIM benchmarks (higher M for higher dimensions)
 */
export declare const PRESET_VERY_HIGH_DIM: Readonly<HNSWPreset>;
/**
 * Preset for small datasets (<10K vectors)
 * Prioritizes recall over speed since brute-force is viable
 */
export declare const PRESET_SMALL_DATASET: Readonly<HNSWPreset>;
/**
 * Preset for large datasets (100K-1M vectors)
 * Balances recall with build time and memory
 *
 * Uses a deeper default search beam than older releases. Large graphs
 * generally need more query-time exploration to keep recall stable.
 */
export declare const PRESET_LARGE_DATASET: Readonly<HNSWPreset>;
/**
 * Preset for maximum recall (prioritizes accuracy over speed)
 * Use when recall is critical and latency is acceptable
 */
export declare const PRESET_MAX_RECALL: Readonly<HNSWPreset>;
/**
 * Preset for minimum latency (prioritizes speed over recall)
 * Use as a latency-oriented starting point when a lower recall target may be
 * acceptable; validate the actual quality/latency trade-off on your workload.
 */
export declare const PRESET_LOW_LATENCY: Readonly<HNSWPreset>;
/**
 * All available presets
 */
export declare const PRESETS: Readonly<Record<string, Readonly<HNSWPreset>>>;
/**
 * Get recommended preset based on dimension and dataset size
 *
 * For high-dimensional vectors (768D+), dimension takes priority over dataset size
 * because recall degrades significantly without higher M values.
 */
export declare function getRecommendedPreset(dimension: number, datasetSize?: number): HNSWPreset;
/**
 * Get a browser-safe preset that caps memory-heavy HNSW parameters.
 *
 * This keeps the high-recall server/runtime defaults intact while giving browser
 * apps a one-call preset for OPFS/WebWorker use, especially on Safari and
 * lower-memory mobile devices where aggressive M/efSearch values can cause
 * long tasks or tab crashes.
 */
export declare function getBrowserRecommendedPreset(dimension: number, options?: BrowserPresetOptions): HNSWPreset;
/**
 * Get preset by name
 */
export declare function getPreset(name: string): HNSWPreset | undefined;
/**
 * RAG-specific preset recommendation
 * For typical RAG applications using popular embedding models
 */
export declare function getRAGPreset(embeddingModel: string): HNSWPreset;
//# sourceMappingURL=presets.d.ts.map