/**
 * Semantic Embedding Index
 *
 * Optional ANN (approximate nearest neighbor) layer on top of the artifact
 * index. Embeds node summaries/titles into dense vectors using a small local
 * model and stores them in an HNSW index for fast similarity queries.
 *
 * Install: npm install @xenova/transformers hnswlib-node
 *
 * @implements #730
 * @source @src/artifacts/types.ts
 * @tests @test/unit/artifacts/embedding-index.test.ts
 */
import type { MetadataEntry } from './types.js';
/**
 * Default embedding model (all-MiniLM-L6-v2: ~22MB, 384 dims, ~5ms/embedding on CPU)
 */
export declare const DEFAULT_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2";
export declare const DEFAULT_EMBEDDING_DIMS = 384;
/**
 * Embedding index manifest stored alongside the HNSW index
 */
export interface EmbeddingManifest {
    /** Model identifier used for embedding */
    model: string;
    /** Vector dimensionality */
    dims: number;
    /** Ordered list of node IDs (position → node ID) */
    nodeIds: string[];
    /** ISO timestamp of last build */
    builtAt: string;
    /** Checksums at build time for incremental detection */
    checksums: Record<string, string>;
}
/**
 * Configuration for the embedding index (from .aiwg/config.yaml)
 */
export interface EmbeddingConfig {
    /** Enable embedding index for this graph */
    enabled: boolean;
    /** Model to use (default: Xenova/all-MiniLM-L6-v2) */
    model?: string;
    /** Number of results for semantic queries */
    topK?: number;
    /** When to rebuild: 'content-change' | 'always' | 'never' */
    rebuildOn?: 'content-change' | 'always' | 'never';
}
/**
 * Semantic search result
 */
export interface SemanticResult {
    /** Node ID (artifact path or REF identifier) */
    nodeId: string;
    /** Cosine similarity score (0-1, higher is more similar) */
    score: number;
}
/**
 * Check if embedding dependencies are available.
 */
export declare function checkEmbeddingDeps(): Promise<{
    available: boolean;
    missing: string[];
}>;
/**
 * Build an embedding index from artifact metadata entries.
 *
 * Embeds each entry's title + summary into a dense vector and stores
 * them in an HNSW index for fast approximate nearest-neighbor queries.
 *
 * @param entries - Map of node ID → MetadataEntry
 * @param outputDir - Directory to write embeddings/ subfolder
 * @param model - Transformer model identifier
 * @returns Number of entries embedded
 */
export declare function buildEmbeddingIndex(entries: Record<string, MetadataEntry>, outputDir: string, model?: string): Promise<number>;
/**
 * Load an embedding manifest from an index directory.
 */
export declare function loadEmbeddingManifest(indexDir: string): EmbeddingManifest | null;
/**
 * Query the embedding index for semantically similar artifacts.
 *
 * @param query - Natural language query string
 * @param indexDir - Directory containing the embeddings/ subfolder
 * @param topK - Number of results to return
 * @returns Ranked list of semantic results
 */
export declare function semanticQuery(query: string, indexDir: string, topK?: number): Promise<SemanticResult[]>;
/**
 * Get semantic neighbors of a specific node.
 *
 * @param nodeId - Node to find neighbors for
 * @param entries - Metadata entries to get the node's text
 * @param indexDir - Directory containing the embeddings/ subfolder
 * @param topK - Number of results
 */
export declare function semanticNeighbors(nodeId: string, entries: Record<string, MetadataEntry>, indexDir: string, topK?: number): Promise<SemanticResult[]>;
/**
 * Determine which entries need re-embedding based on checksum changes.
 *
 * @param entries - Current metadata entries
 * @param manifest - Existing embedding manifest
 * @returns Object with entries that changed and entries that are new
 */
export declare function detectEmbeddingChanges(entries: Record<string, MetadataEntry>, manifest: EmbeddingManifest): {
    changed: string[];
    added: string[];
    removed: string[];
};
/** A near-duplicate pair surfaced by the embedding index. */
export interface DedupPair {
    /** First node id (lexicographically smaller). */
    a: string;
    /** Second node id. */
    b: string;
    /** Cosine similarity (0-1). */
    score: number;
}
/**
 * Near-duplicate report (#1493). Reads the prebuilt embedding index and, using
 * each stored vector (no re-embedding), finds node pairs whose cosine
 * similarity is at or above `threshold`. Pairs are de-duplicated (unordered)
 * and returned most-similar-first.
 *
 * Requires a built embedding index (`aiwg index embed`). Throws if absent.
 */
export declare function dedupReport(indexDir: string, threshold?: number, topK?: number): Promise<DedupPair[]>;
//# sourceMappingURL=embedding-index.d.ts.map