ai-rpg/backend/app/config.py

"""Application configuration loaded from environment + DB-backed admin settings."""
from __future__ import annotations

import os
import secrets
from functools import lru_cache
from typing import Annotated, List

from pydantic import Field, field_validator
from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict


class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env", extra="ignore", case_sensitive=False)

    # Database
    database_url: str = "postgresql+asyncpg://airpg:airpg_secret@localhost:5432/airpg"

    # Redis
    redis_url: str = "redis://localhost:6379/0"

    # Qdrant
    qdrant_url: str = "http://localhost:6333"

    # Auth
    jwt_secret: str = Field(default_factory=lambda: secrets.token_hex(32))
    jwt_algorithm: str = "HS256"
    access_token_expire_minutes: int = 60 * 24 * 7  # 7 days

    # Admin setup
    # If empty, will be generated at first run and printed to console.
    admin_setup_token: str = ""

    # CORS
    # `NoDecode` tells pydantic-settings NOT to JSON-parse the env value, so a
    # plain comma-separated string like "http://localhost:5173,http://localhost:8080"
    # reaches our `@field_validator` intact, which then splits it into a list.
    cors_origins: Annotated[List[str], NoDecode] = Field(
        default_factory=lambda: ["http://localhost:5173"]
    )

    @field_validator("cors_origins", mode="before")
    @classmethod
    def _split_origins(cls, v):
        if isinstance(v, str):
            # Allow both JSON arrays (e.g. '["http://a","http://b"]') and plain
            # comma-separated strings (e.g. "http://a,http://b") from env vars.
            v = v.strip()
            if v.startswith("["):
                import json
                try:
                    return [o.strip() for o in json.loads(v) if o.strip()]
                except Exception:
                    pass
            return [o.strip() for o in v.split(",") if o.strip()]
        if isinstance(v, (list, tuple)):
            return [str(o).strip() for o in v if str(o).strip()]
        return v

    # Logging
    log_level: str = "INFO"

    # Default LLM (used to seed DB on first run; overridable via admin panel)
    default_llm_base_url: str = "http://localhost:1234/v1"
    default_llm_api_key: str = "dummy"
    default_llm_model: str = "local-model"

    # Default embeddings / RAG settings (overridable via admin panel)
    # provider="hash" is a deterministic offline fallback (no semantic quality).
    # Switch to "openai" and point embedding.base_url at an OpenAI-compatible /embeddings endpoint
    # for real semantic search.
    default_embedding_provider: str = "hash"
    default_embedding_base_url: str = ""  # empty = reuse llm.base_url
    default_embedding_api_key: str = ""   # empty = reuse llm.api_key
    default_embedding_model: str = "text-embedding-3-small"
    default_embedding_dim: int = 0        # 0 = auto-probe from endpoint
    default_embedding_request_timeout: int = 60

    # Context manager defaults (admin-overridable)
    default_recent_messages: int = 10
    default_compress_threshold: int = 20
    default_summary_messages: int = 10

    # Worker mode flag
    worker_mode: bool = False

    @property
    def is_worker(self) -> bool:
        return bool(os.getenv("WORKER_MODE")) or self.worker_mode


@lru_cache
def get_settings() -> Settings:
    return Settings()


settings = get_settings()