/**
 * Embedding index store — in-memory HNSW-like nearest-neighbor search.
 *
 * Stores embedding vectors alongside chunk metadata for semantic retrieval.
 * Uses brute-force cosine similarity for simplicity and correctness.
 * For large codebases, a future version can swap in SQLite vss or HNSW.
 *
 * Design: all in-memory, rebuilt on startup from indexed chunks.
 */
import type { CodeChunk } from './chunker';
import type { EmbeddingProvider } from './provider';
export interface IndexedChunk {
    chunk: CodeChunk;
    embedding: Float32Array;
}
export interface SearchResult {
    chunk: CodeChunk;
    score: number;
}
export interface IndexStore {
    /** Number of indexed chunks. */
    readonly size: number;
    /** Add chunks with their embeddings. */
    add(items: IndexedChunk[]): void;
    /** Remove all chunks for a given file path. */
    removeFile(filePath: string): void;
    /** Search for the top-k most similar chunks to a query embedding. */
    search(queryEmbedding: Float32Array, topK?: number): SearchResult[];
    /** Clear the entire index. */
    clear(): void;
    /** Get all indexed file paths. */
    indexedFiles(): string[];
    /**
     * Optional BM25 keyword search. Stores that do not support keyword search
     * omit this method; callers must check before invoking.
     */
    keywordSearch?(query: string, topK?: number): SearchResult[];
}
/**
 * Create an in-memory index store.
 */
export declare function createIndexStore(): IndexStore;
export interface IndexingResult {
    chunksIndexed: number;
    filesIndexed: number;
    durationMs: number;
}
/**
 * Index a set of code chunks using the given embedding provider.
 * Processes in batches for efficiency.
 */
export declare function indexChunks(chunks: CodeChunk[], provider: EmbeddingProvider, store: IndexStore, batchSize?: number): Promise<IndexingResult>;
/**
 * Search the index with a natural language query.
 *
 * Modes:
 *   - `semantic` (default): cosine similarity over dense embeddings.
 *   - `keyword`: BM25 via the store's `keywordSearch` (FTS5). Requires a store
 *     that implements `keywordSearch` (e.g., `createSqliteIndexStore`).
 *   - `hybrid`: fuse semantic + keyword rankings with Reciprocal Rank Fusion.
 *     Falls back to `semantic` if the store has no `keywordSearch`.
 */
export declare function semanticSearch(query: string, provider: EmbeddingProvider, store: IndexStore, topK?: number, mode?: 'semantic' | 'keyword' | 'hybrid'): Promise<SearchResult[]>;
//# sourceMappingURL=index-store.d.ts.map