""" 全局配置中心 —— 从环境变量 / .env 文件加载所有配置项，提供开发环境默认值。 Application settings loaded from environment variables with sensible defaults. """ from __future__ import annotations from pathlib import Path from pydantic_settings import BaseSettings, SettingsConfigDict class Settings(BaseSettings): """后端全局配置类，涵盖 ES、Neo4j、Redis、LLM、Embedding 等所有基础设施连接参数及业务开关。 Central configuration for the zm-rag backend. Values are read from environment variables (case-insensitive) or a `.env` file located next to this module. Every field has a default that matches the docker-compose development environment. """ # 使用基于 __file__ 的绝对路径定位 .env，避免工作目录不同导致加载失败 # Use absolute path based on __file__ to locate .env reliably model_config = SettingsConfigDict( env_file=str(Path(__file__).resolve().parent.parent / ".env"), env_file_encoding="utf-8", case_sensitive=False, extra="ignore", ) # ── Application ────────────────────────────────────────────────────── app_name: str = "zm-rag" app_version: str = "0.1.0" debug: bool = False # ── Elasticsearch ──────────────────────────────────────────────────── es_host: str = "http://localhost:9200" es_username: str = "" es_password: str = "" es_chunk_index: str = "gov_doc_chunks" es_meta_index: str = "gov_doc_meta" es_service_guide_index: str = "gov_service_guides" es_request_timeout: int = 30 # TLS 证书验证开关，生产环境建议设为 True 以防止中间人攻击 es_verify_certs: bool = False # ES 副本数，生产环境建议设为 1 或更高以保证数据高可用 es_number_of_replicas: int = 0 # ── Neo4j ──────────────────────────────────────────────────────────── neo4j_uri: str = "bolt://localhost:7687" neo4j_user: str = "neo4j" neo4j_password: str = "zm_rag_2024" neo4j_database: str = "neo4j" # ── Redis ──────────────────────────────────────────────────────────── redis_url: str = "redis://localhost:6379/0" redis_permissions_ttl: int = 300 # seconds # ── MySQL ──────────────────────────────────────────────────────────── mysql_enabled: bool = False mysql_host: str = "" mysql_port: int = 3306 mysql_user: str = "" mysql_password: str = "" mysql_database: str = "" mysql_charset: str = "utf8mb4" mysql_connect_timeout: int = 5 mysql_pool_minsize: int = 1 mysql_pool_maxsize: int = 10 # ── Research session storage ──────────────────────────────────────── research_session_backend: str = "mysql" research_session_ttl_seconds: int = 604800 # ── Celery ─────────────────────────────────────────────────────────── celery_broker_url: str = "redis://localhost:6379/1" celery_result_backend: str = "redis://localhost:6379/2" # ── LLM (OpenAI-compatible) ────────────────────────────────────────── llm_base_url: str = "http://localhost:11434/v1" # 生产环境必须配置真实的 API Key，"no-key" 仅用于本地开发（如 Ollama） # MUST be configured with a real API key in production; "no-key" is for local dev only (e.g. Ollama) llm_api_key: str = "no-key" llm_model: str = "qwen2.5:14b" llm_temperature: float = 0.3 llm_max_tokens: int = 4096 # Set to False to disable chain-of-thought / thinking (e.g. DashScope Qwen) llm_enable_thinking: bool = True # ── Embedding (OpenAI-compatible) ──────────────────────────────────── embedding_base_url: str = "http://localhost:11434/v1" embedding_api_key: str = "no-key" embedding_model: str = "bge-m3:latest" embedding_dimensions: int = 1024 # Max texts per batch for embedding API calls. # DashScope text-embedding-v3 limit is 10; set lower to be conservative. embedding_batch_size: int = 6 # ── JWT / Auth ─────────────────────────────────────────────────────── jwt_secret: str = "zm-rag-dev-secret-change-in-production" jwt_algorithm: str = "HS256" # Service-to-service token for the Big Lobster KB-compatible adapter. # Configure LOBSTER_KB_API_KEY in production; the adapter is unavailable # when this value is empty. lobster_kb_api_key: str = "" lobster_kb_detail_max_chars: int = 40_000 # ── Data paths ─────────────────────────────────────────────────────── # Optional service-to-service token for /ingest/webhook/*. # When empty, webhook ingest remains unauthenticated for local/dev # compatibility. When configured, callers must send: # Authorization: Bearer rag_ingest_token: str = "" file_storage_path: Path = Path("/data/files") @property def pdf_storage_path(self) -> Path: """Alias for backward compatibility.""" return self.file_storage_path # ── Multi-format document support ───────────────────────────────── # 使用 tuple 作为默认值，避免可变默认值在实例间共享导致意外修改 # Use tuple as default to prevent mutable default sharing across instances supported_file_types: tuple[str, ...] = ( "pdf", "doc", "docx", "xls", "xlsx", "ppt", "pptx", "wps", "et", "ofd", "png", "jpg", "jpeg", "tiff", "bmp", "txt", "md", "markdown", ) max_file_size_mb: int = 100 # Docling docling_ocr_enabled: bool = True docling_max_tokens: int = 512 # Java document converter service converter_base_url: str = "http://localhost:18800" converter_timeout: int = 120 # Formats that need conversion via Java service before Docling processing formats_need_conversion: list[str] = ["doc", "xls", "ppt", "wps", "et", "ofd"] # ── Knowledge Graph ─────────────────────────────────────────────────── # Set to False to skip graph construction during document ingest. graph_build_enabled: bool = True # Maximum body characters forwarded to the LLM for entity extraction. graph_max_content_chars: int = 12_000 # ── Summary Generation ──────────────────────────────────────────────── # Set to False to disable automatic summary generation during ingest. summary_enabled: bool = True # Maximum document characters forwarded to the LLM for summarisation. summary_max_content_chars: int = 8_000 # LLM 摘要生成的最大 token 数 / Max tokens for LLM summary generation summary_max_tokens: int = 512 # ── Document Analysis (unified metadata + summary) ──────────────────── # 统一文档分析：将元数据提取和摘要生成合并为单次 LLM 调用 # Max document characters for the unified metadata+summary LLM call. analysis_max_content_chars: int = 50_000 # ── Rate Limiting ───────────────────────────────────────────────────── rate_limit_enabled: bool = True rate_limit_research_max: int = 10 # max requests per window rate_limit_research_window: int = 60 # window in seconds rate_limit_search_max: int = 60 rate_limit_search_window: int = 60 rate_limit_default_max: int = 200 rate_limit_default_window: int = 60 # ── Cache TTLs ──────────────────────────────────────────────────────── # Set to 0 to disable embedding cache. embedding_cache_ttl: int = 3600 # seconds search_cache_ttl: int = 120 # seconds (0 = disabled) # 检索结果分数阈值，低于此值的结果将被过滤（0 = 不过滤） search_score_threshold: float = 0.5 # ── Ingest Tracing ────────────────────────────────────────────────── es_trace_index: str = "gov_doc_ingest_traces" es_event_index: str = "gov_doc_ingest_events" es_artifact_index: str = "gov_doc_ingest_artifacts" # Retention days for cleanup tasks trace_retention_days: int = 180 event_retention_days: int = 90 artifact_retention_days_standard: int = 30 artifact_retention_days_debug: int = 7 # ── CORS ───────────────────────────────────────────────────────────── cors_origins: list[str] = ["http://localhost:3000", "http://localhost:5173"] # Singleton – import this wherever settings are needed. settings = Settings()