vkus-ai-v4/backend/app/config.py

"""Загрузка llm.config.yaml и переопределений из переменных окружения."""

from __future__ import annotations

import os
from functools import lru_cache
from pathlib import Path
from typing import Any

import yaml
from dotenv import load_dotenv

PROJECT_ROOT = Path(__file__).resolve().parents[2]
CONFIG_PATH = PROJECT_ROOT / "llm.config.yaml"

# .env в корне репозитория (секреты не коммитить)
load_dotenv(PROJECT_ROOT / ".env")


def _deep_get(data: dict[str, Any], *keys: str, default: Any = None) -> Any:
    node: Any = data
    for key in keys:
        if not isinstance(node, dict):
            return default
        node = node.get(key)
        if node is None:
            return default
    return node


@lru_cache(maxsize=1)
def load_config() -> dict[str, Any]:
    if not CONFIG_PATH.exists():
        return {}
    with CONFIG_PATH.open(encoding="utf-8") as fh:
        return yaml.safe_load(fh) or {}


class Settings:
  """Настройки интеграции LLM и временного локального RAG."""

  @property
  def project_root(self) -> Path:
      return PROJECT_ROOT

  @property
  def llm_enabled(self) -> bool:
      env = os.getenv("VKUS_LLM_ENABLED")
      if env is not None:
          return env.lower() in ("1", "true", "yes")
      return bool(_deep_get(load_config(), "provider", "enabled", default=False))

  @property
  def llm_base_url(self) -> str:
      # Полный URL из .env (OpenRouter и др.) — нормализуем до base
      api_url = os.getenv("VKUS_LLM_API_URL")
      if api_url:
          url = api_url.rstrip("/")
          if url.endswith("/v1/chat/completions"):
              return url[: -len("/v1/chat/completions")]
          return url
      return os.getenv(
          "VKUS_LLM_BASE_URL",
          _deep_get(load_config(), "api", "base_url", default="https://apiai.sibur.local"),
      )

  @property
  def llm_chat_url(self) -> str:
      """Полный URL для chat/completions."""
      api_url = os.getenv("VKUS_LLM_API_URL")
      if api_url:
          return api_url.rstrip("/")
      return f"{self.llm_base_url.rstrip('/')}/v1/chat/completions"

  @property
  def llm_model(self) -> str:
      return os.getenv(
          "VKUS_LLM_MODEL",
          _deep_get(load_config(), "llm", "model", default="gpt-oss-120b"),
      )

  @property
  def llm_api_key(self) -> str | None:
      key_name = _deep_get(load_config(), "auth", "api_key_env", default="VKUS_LLM_API_KEY")
      return os.getenv(key_name or "VKUS_LLM_API_KEY") or None

  @property
  def llm_jwt_token(self) -> str | None:
      key_name = _deep_get(load_config(), "auth", "jwt_token_env", default="VKUS_LLM_JWT_TOKEN")
      return os.getenv(key_name or "VKUS_LLM_JWT_TOKEN") or None

  @property
  def llm_timeout(self) -> float:
      env = os.getenv("VKUS_LLM_TIMEOUT_SEC")
      if env:
          return float(env)
      return float(_deep_get(load_config(), "api", "timeout_seconds", default=60))

  @property
  def llm_verify_ssl(self) -> bool:
      return bool(_deep_get(load_config(), "api", "verify_ssl", default=True))

  @property
  def llm_temperature(self) -> float:
      env = os.getenv("VKUS_LLM_TEMPERATURE")
      if env:
          return float(env)
      return float(_deep_get(load_config(), "llm", "parameters", "temperature", default=0.3))

  @property
  def llm_max_tokens(self) -> int:
      env = os.getenv("VKUS_LLM_MAX_TOKENS")
      if env:
          return int(env)
      return int(_deep_get(load_config(), "llm", "parameters", "max_tokens", default=2048))

  @property
  def system_prompt_path(self) -> Path:
      rel = _deep_get(
          load_config(),
          "llm",
          "system_prompt",
          "file",
          default="Start data/Системный промпт - правила общения с пользователем.md",
      )
      return PROJECT_ROOT / str(rel)

  @property
  def system_prompt_fallback(self) -> str:
      return str(
          _deep_get(
              load_config(),
              "llm",
              "system_prompt",
              "fallback",
              default="Ты — корпоративный ИТ-ассистент портала ВКУС.",
          )
      )

  @property
  def confidence_threshold(self) -> int:
      return int(_deep_get(load_config(), "assistant", "confidence_threshold", default=70))

  # --- Временный локальный RAG (внутри backend) ---

  @property
  def rag_mode(self) -> str:
      """local — временный RAG в backend; external — целевая платформа ИИ (будущее)."""
      return str(_deep_get(load_config(), "rag", "mode", default="local"))

  @property
  def rag_local_enabled(self) -> bool:
      env = os.getenv("VKUS_RAG_LOCAL_ENABLED")
      if env is not None:
          return env.lower() in ("1", "true", "yes")
      return bool(_deep_get(load_config(), "rag", "local", "enabled", default=True))

  @property
  def rag_documents_path(self) -> Path:
      rel = _deep_get(load_config(), "rag", "local", "documents_path", default="Manuals")
      return PROJECT_ROOT / str(rel)

  @property
  def rag_top_k(self) -> int:
      return int(_deep_get(load_config(), "rag", "local", "top_k", default=5))

  @property
  def rag_min_score(self) -> float:
      return float(_deep_get(load_config(), "rag", "local", "min_score", default=0.12))

  @property
  def rag_chunk_size(self) -> int:
      return int(_deep_get(load_config(), "rag", "local", "chunk_size", default=450))

  @property
  def use_pattern_kb(self) -> bool:
      return bool(_deep_get(load_config(), "fallback", "use_local_knowledge_base", default=True))

  @property
  def use_template_on_llm_failure(self) -> bool:
      return bool(_deep_get(load_config(), "fallback", "use_template_on_llm_failure", default=True))


settings = Settings()