import type { LocalModelTier } from '../../integrations/cloud/llm/local-tier.js';
export declare const MAX_MARKDOWN_CHARS = 6000;
export interface LocalLlmRequest {
    schema: Record<string, unknown>;
    html: string;
    url: string;
    /** Resolved local-model tier (endpoint + model) from resolveLocalModelTier. */
    tier: LocalModelTier;
}
/**
 * Ask the local model to fill a schema from the DETERMINISTIC pre-extraction of
 * a page (structured brief + trimmed markdown) rather than raw HTML. The result
 * is parsed + validated against the schema by runLlmJson. On any failure —
 * timeout, non-200, invalid JSON, transport error — this returns `null` so the
 * caller falls back to the deterministic path. Never throws.
 *
 * The resolved tier's endpoint/model are threaded to runLlmJson via the
 * per-call `backend` override — NOT process.env. That keeps the call routed to
 * the local server without ever mutating ambient env, so concurrent extract
 * calls can never corrupt a shared WIGOLO_LLM_PROVIDER (which would silently
 * reroute cloud→local for every other subsystem for the rest of the process).
 */
export declare function extractWithLocalLlm(request: LocalLlmRequest): Promise<Record<string, unknown> | null>;
//# sourceMappingURL=local-llm.d.ts.map