/**
 * Persistence for file-level semantic embeddings (`file_vectors`).
 *
 * Vectors from two different models are not comparable — the spaces have
 * nothing to do with each other — so the provider id is stored alongside every
 * row and a provider change wipes the table rather than silently mixing them.
 */
import type { DatabaseSync } from 'node:sqlite';
type Statement = ReturnType<DatabaseSync['prepare']>;
type PrepareStatement = (sql: string) => Statement;
/** Metadata key recording which provider produced the stored vectors. */
export declare const FILE_VECTOR_PROVIDER_KEY = "file_vector_provider";
export interface FileVectorRow {
    file: string;
    vector: Float32Array;
    /** Hash of the exact text that was embedded. */
    sourceHash: string;
    provider: string;
}
/** What is already embedded, so a pass can skip unchanged text. */
export interface FileVectorState {
    file: string;
    sourceHash: string;
}
export declare function getFileVectorStatesWithStatement(stmt: PrepareStatement, provider: string): Map<string, string>;
export declare function upsertFileVectorsWithStatement(stmt: PrepareStatement, maxSqlVars: number, rows: readonly FileVectorRow[]): void;
/**
 * Drop every stored vector when the provider changes.
 *
 * Returns true when a wipe happened, so the caller can report that the next
 * pass is a full re-embed rather than an incremental one.
 */
export declare function reconcileVectorProviderWithStatement(stmt: PrepareStatement, getMetadata: (key: string) => string | undefined, setMetadata: (key: string, value: string) => void, provider: string): boolean;
/** Vectors dropped for files no longer indexed. */
export declare function pruneOrphanFileVectorsWithStatement(stmt: PrepareStatement): number;
export declare function countFileVectorsWithStatement(stmt: PrepareStatement): number;
export interface VectorHit {
    file: string;
    /** Cosine similarity in [-1, 1]. */
    score: number;
}
/**
 * Rank every stored vector against `query` and return the closest files.
 *
 * A brute-force scan, deliberately: at file granularity this repository has
 * roughly eight thousand vectors of 384 floats, which is about twelve
 * megabytes and a few milliseconds to sweep. An approximate-nearest-neighbour
 * index would add a dependency and an accuracy cliff to save time nobody is
 * waiting on.
 */
export declare function searchFileVectorsWithStatement(stmt: PrepareStatement, query: Float32Array, limit: number, minScore: number): VectorHit[];
export {};
//# sourceMappingURL=writer-vectors.d.ts.map