import type { Embedder } from './types.js';
/**
 * Build the Embedder for the resolved (provider, model) pair.
 *
 * Single chokepoint: `resolvePresetConfig(process.env)` returns provider AND
 * model atomically — there is no path through this function where the two
 * can desync. (Pre-v1.7.8 the Ollama branch read EMBEDDING_MODEL on its own
 * and ignored EMBEDDING_PRESET, which is how `multilingual-ollama` silently
 * fell through to `nomic-embed-text` for users who set the preset.)
 *
 * Returns an Embedder that still needs `init()` called on it (same contract
 * as `new TransformersEmbedder()`); the caller's existing
 * ensureEmbedderReady/init sequence handles both providers uniformly.
 * For the Ollama path, `init()` probes the server once to populate
 * `dimensions()`, unless `OLLAMA_EMBEDDING_DIM` declared it up front.
 *
 * Note on `TransformersEmbedder`'s own DEFAULT_MODEL (`Xenova/all-MiniLM-L6-v2`):
 * intentionally left as MiniLM so tests that construct it directly (without
 * going through this factory) stay deterministic and don't depend on
 * process.env state. Production paths always come through here.
 */
export declare function createEmbedder(): Embedder;