scrypt/inference/__init__.py

"""Backend selection: local llama-server > API > scripted, by availability.

Environment knobs:
  SCRYPT_BACKEND   auto|local|api|scripted  (default auto)
  SCRYPT_API_BASE  OpenAI-compatible base URL (default OpenRouter)
  SCRYPT_API_KEY   bearer token for the API
  SCRYPT_MODEL     model id for the API
"""

from __future__ import annotations

import os

from .api import DEFAULT_MODEL, OpenAIChatBackend
from .backend import Backend, ScriptedBackend, complete
from .local import LlamaServer, preflight

OPENROUTER = "https://openrouter.ai/api/v1"


def build_backend() -> tuple[Backend, LlamaServer | None, str]:
    """Returns (backend, server-or-None, mode-name)."""
    mode = os.environ.get("SCRYPT_BACKEND", "auto")
    api_key = os.environ.get("SCRYPT_API_KEY", "")
    api_base = os.environ.get("SCRYPT_API_BASE", OPENROUTER)
    model = os.environ.get("SCRYPT_MODEL", DEFAULT_MODEL)

    if mode in ("auto", "local") and not preflight():
        server = LlamaServer()
        server.start()
        return server.backend(), server, "local"
    if mode == "local":
        raise RuntimeError("SCRYPT_BACKEND=local but: " + "; ".join(preflight()))
    if mode in ("auto", "api") and api_key:
        return OpenAIChatBackend(api_base, model=model, api_key=api_key), None, "api"
    if mode == "api":
        raise RuntimeError("SCRYPT_BACKEND=api requires SCRYPT_API_KEY")
    return ScriptedBackend(), None, "scripted"


__all__ = [
    "Backend",
    "OpenAIChatBackend",
    "ScriptedBackend",
    "build_backend",
    "complete",
]