/**
 * Semantic side of monograph search, as seen by the MCP tools: the local
 * embedding model, the background job that fills the vector index, and the
 * decision of when a query may use it. Everything degrades to lexical search —
 * no model, no vectors yet, or `MONOGRAPH_EMBED=0` — never to an error.
 */
import type { Embedder } from '@monoes/monograph';
/** The local sentence-embedding model, loaded once per process; null if unavailable. */
export declare function getEmbedder(): Promise<Embedder | null>;
/** Why semantic search is off, for messages: no model on disk (or switched off) vs. just not built yet. */
export declare function semanticModelAvailable(): Promise<boolean>;
/**
 * Fire-and-forget: embed whatever summaries lack a current vector. Runs in
 * this (long-lived) process behind a guard so repeated tool calls and builds
 * don't stack jobs; resumable, so an interrupted run loses at most one batch.
 */
export declare function triggerBackgroundEmbed(dbPath: string): boolean;
/**
 * The embedder to search with, or null when semantic search can't help yet:
 * nothing embedded (the job is started once per process and lexical search
 * answers meanwhile) or no model on this machine.
 */
export declare function semanticEmbedder(db: Parameters<typeof import('@monoes/monograph').embeddingStatus>[0], dbPath: string): Promise<Embedder | null>;
//# sourceMappingURL=semantic.d.ts.map