"""Model loaders — local GGUF LLM client (llama.cpp) + setup/download helpers.

Lazy by design: importing this package does NOT import llama_cpp or load any model — that happens only on
first ``generate``/``ensure``. So the gates run without a model downloaded; the rank/JD path is offline once
the weights are cached under ``data/artifacts/models/``.
"""

from .llm import LlmClient, index_gen_client
from .prefetch import ensure_all_models

__all__ = ["LlmClient", "ensure_all_models", "index_gen_client"]