export declare function initEmbedder(): Promise<void>;
export declare function embed(texts: string[]): Promise<Float32Array[]>;
export declare function cosineSimilarity(a: Float32Array, b: Float32Array): number;
/**
 * Per-token embedding for late-interaction reranking (#29).
 *
 * Returns one TokenEmbedding per input text, where each TokenEmbedding holds
 * the L2-normalized per-token hidden states in a flat Float32Array of length
 * `len * dim`. The flat layout (vs `Float32Array[]`) is intentional: the
 * MaxSim hot path benchmarks 3-5x faster on a contiguous buffer because V8
 * inlines the tight `q · d` loop aggressively.
 *
 * Padding (attention_mask = 0) and special tokens (CLS at position 0, SEP at
 * position word-count + 1) are stripped — only "real content" tokens come
 * back. This matches what ColBERT-style scoring expects: the scoring runs
 * over genuinely meaningful positions, not padding.
 *
 * Returns null when the ONNX runtime is unavailable. Callers (rerank.ts) use
 * that to silently fall through to pass-through behavior — never throw, the
 * production retrieval path must not break behind a model load failure.
 *
 * Shape contract:
 *   embedTokens(["hello world"]) → [{ tokens: Float32Array([...]), dim: 384, len: 2 }]
 *   tokens[t * dim + d] is the d-th component of the t-th token's vector.
 */
export interface TokenEmbedding {
    /** Flat per-token hidden states of length len * dim. L2-normalized per token. */
    tokens: Float32Array;
    /** Embedding dimensionality (384 for all-MiniLM-L6-v2). */
    dim: number;
    /** Number of real tokens kept (CLS/SEP/padding stripped). */
    len: number;
}
export declare function embedTokens(texts: string[]): Promise<(TokenEmbedding | null)[]>;
export declare const EMBEDDING_DIM = 384;