/**
 * embedding-config.ts — one place that decides which embedding model the
 * memory path loads and what dimension its vectors have (#2305).
 *
 * `embeddings init` (commands/embeddings.ts) and the `embeddings_init` MCP
 * tool write `<project>/.claude-flow/embeddings.json` with `{ model,
 * dimension, ... }`. Before this module only those two files read it: the
 * memory path, the AgentDB bridge and the router's task embedder all loaded
 * 'Xenova/all-MiniLM-L6-v2' and assumed 384 dimensions regardless of what the
 * user configured.
 *
 * Precedence: `.claude-flow/embeddings.json` → defaults (MiniLM / 384). There
 * is no embedding-model environment override in this codebase today, so none
 * is invented here.
 *
 * Safety: a store that already holds vectors keeps its recorded dimension.
 * When the configured dimension differs, `resolveEffectiveEmbeddingConfig()`
 * returns the recorded shape and warns once — it never mixes dimensions in one
 * store and never rebuilds or rewrites existing vectors.
 */
export declare const DEFAULT_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2";
export declare const DEFAULT_EMBEDDING_DIMENSION = 384;
/** Same directory + file name commands/embeddings.ts and embeddings-tools.ts write. */
export declare const EMBEDDINGS_CONFIG_DIR = ".claude-flow";
export declare const EMBEDDINGS_CONFIG_FILE = "embeddings.json";
export interface EmbeddingConfig {
    /** Model id passed to the transformers `pipeline()` (always namespaced, e.g. `Xenova/...`). */
    model: string;
    dimension: number;
    /** `config` = read from embeddings.json; `recorded` = kept from an existing store; `default` otherwise. */
    source: 'default' | 'config' | 'recorded';
    configPath: string;
}
/** Shape of the vectors an existing store already holds. */
export interface RecordedEmbeddingShape {
    dimension: number;
    model?: string | null;
}
export type RecordedShapeReader = (dbPath: string) => Promise<RecordedEmbeddingShape | null>;
/**
 * `embeddings init` records the bare name it was given (default
 * `all-MiniLM-L6-v2`), while the MCP tool records `Xenova/...`. The
 * transformers pipeline needs the namespaced id, so bare names get `Xenova/`.
 */
export declare function normalizeEmbeddingModel(model: string): string;
/**
 * Resolve the configured embedding model + dimension for a project root
 * (default: cwd, which is where `embeddings init` writes its config).
 * Cached per root; never throws.
 */
export declare function resolveEmbeddingConfig(root?: string): EmbeddingConfig;
/**
 * Keep an existing store's recorded dimension when it differs from the
 * configured one. Warns once per store; returns the configured value
 * untouched when there is nothing recorded or the dimensions agree.
 */
export declare function reconcileWithRecordedShape(configured: EmbeddingConfig, recorded: RecordedEmbeddingShape | null, storeLabel: string): EmbeddingConfig;
/**
 * Default reader: the most recently written real (non-hash) vector in
 * `memory_entries`. A store with no such vector records nothing, so a fresh
 * or empty database follows the configuration.
 */
export declare const readRecordedEmbeddingShape: RecordedShapeReader;
/**
 * The embedding config to actually use: the configured model, unless one of
 * the stores in `dbPaths` already holds vectors of another dimension. The
 * memory path writes the same vectors to memory.db and (through the AgentDB
 * bridge) agentdb-memory.db, so callers pass both and every caller converges
 * on one answer: the first store with recorded vectors wins. Cached per
 * (root, dbPaths).
 */
export declare function resolveEffectiveEmbeddingConfig(options?: {
    root?: string;
    dbPaths?: ReadonlyArray<string | undefined>;
    readRecordedShape?: RecordedShapeReader;
}): Promise<EmbeddingConfig>;
/** Test seam: forget cached resolutions and emitted warnings. */
export declare function _resetEmbeddingConfigCache(): void;
//# sourceMappingURL=embedding-config.d.ts.map