src/config.py

import os
from pathlib import Path

from dotenv import load_dotenv
from pydantic import Field
from pydantic_settings import BaseSettings

load_dotenv()

PROJECT_ROOT = Path(__file__).resolve().parent.parent

DATA_DIR = Path(os.getenv("DATA_DIR", PROJECT_ROOT / "data"))
DATA_INPUT_DIR = Path(os.getenv("DATA_INPUT_DIR", PROJECT_ROOT / "test_data"))
DATA_OUTPUT_DIR = Path(os.getenv("DATA_OUTPUT_DIR", PROJECT_ROOT / "output"))
DATA_CRAWLED_DIR = Path(os.getenv("DATA_CRAWLED_DIR", DATA_DIR / "crawl"))
BATCH_SIZE = 1


class Settings(BaseSettings):
    """Application settings with environment variable support."""

    # MegaLLM API settings (for small model)
    megallm_api_key: str = Field(
        default="",
        alias="MEGALLM_API_KEY",
        description="API key for MegaLLM",
    )
    megallm_base_url: str = Field(
        default="https://ai.megallm.io/v1",
        alias="MEGALLM_BASE_URL",
    )
    
    # Groq API settings (for large model)
    groq_api_key: str = Field(
        default="",
        alias="GROQ_API_KEY",
        description="API key for Groq",
    )
    groq_base_url: str = Field(
        default="https://api.groq.com/openai/v1",
        alias="GROQ_BASE_URL",
    )
    
    # OpenRouter API (fallback)
    openrouter_api_key: str = Field(
        default="",
        alias="OPENROUTER_API_KEY",
        description="API key for OpenRouter (fallback)",
    )
    
    # Model names
    model_small: str = Field(
        default="qwen/qwen3-32b",
        alias="MODEL_SMALL",
        description="Small model for routing, reranking, and RAG",
    )
    model_large: str = Field(
        default="meta-llama/llama-4-scout-17b-16e-instruct",
        alias="MODEL_LARGE",
        description="Large model for logic/direct answering",
    )
    
    # Available large models for testing
    available_large_models: list[str] = [
        "llama-3.3-70b-versatile",
        "meta-llama/llama-4-scout-17b-16e-instruct", 
        "moonshotai/kimi-k2-instruct-0905",
        "openai/gpt-oss-120b"
    ]

    # Local embedding model (Vietnamese)
    embedding_model: str = Field(
        default="bkai-foundation-models/vietnamese-bi-encoder",
        alias="EMBEDDING_MODEL",
    )

    # Vector database
    qdrant_collection: str = Field(
        default="vnpt_knowledge_base",
        alias="QDRANT_COLLECTION",
    )
    vector_db_path: str = Field(
        default="",
        alias="VECTOR_DB_PATH",
        description="Path to Qdrant storage. Defaults to DATA_DIR/qdrant_storage if empty.",
    )
    
    # Firebase Admin
    firebase_service_account_path: str = Field(
        default="serviceAccountKey.json",
        alias="FIREBASE_SERVICE_ACCOUNT_PATH",
        description="Path to Firebase Service Account JSON",
    )
    
    firebase_credentials_json: str = Field(
        default="",
        alias="FIREBASE_CREDENTIALS_JSON",
        description="Raw JSON string of service account key (for Cloud/HF Env)",
    )

    chunk_size: int = 1000
    chunk_overlap: int = 200
    top_k_retrieval: int = 10
    top_k_rerank: int = 3

    @property
    def vector_db_path_resolved(self) -> Path:
        """Resolve vector database path, defaulting to DATA_DIR/qdrant_storage."""
        if self.vector_db_path:
            return Path(self.vector_db_path)
        return DATA_DIR / "qdrant_storage"

    class Config:
        env_file = ".env"
        extra = "ignore"


settings = Settings()

# Validate API key on import
if not settings.megallm_api_key:
    import warnings
    warnings.warn("MEGALLM_API_KEY not set. LLM calls will fail.")