Initial commit: InternalAuditInterprise

backend/app/llm/__init__.py

@@ -0,0 +1,10 @@
+"""LLM Provider 抽象层。
+
+通过统一接口隔离 LLM 实现，使开发期可用公网千问、生产期无缝切换本地 vLLM。
+强约束："数据零出域"红线由 provider 工厂在 prod 环境拦截公网 Provider。
+"""
+
+from app.llm.base import ChatMessage, LLMProvider, LLMResponse
+from app.llm.factory import get_llm_provider
+
+__all__ = ["ChatMessage", "LLMProvider", "LLMResponse", "get_llm_provider"]

backend/app/llm/base.py

@@ -0,0 +1,44 @@
+"""LLM Provider 抽象接口与数据模型。"""
+
+from __future__ import annotations
+
+import abc
+from dataclasses import dataclass, field
+
+
+@dataclass
+class ChatMessage:
+    role: str  # "system" | "user" | "assistant"
+    content: str
+
+
+@dataclass
+class LLMResponse:
+    content: str
+    model: str
+    provider: str
+    # 是否经过出域（公网）通道，便于审计轨迹记录
+    egress: bool = False
+    raw: dict = field(default_factory=dict)
+
+
+class LLMProvider(abc.ABC):
+    """所有 LLM 实现的统一接口。
+
+    业务代码只依赖本接口；切换公网/本地仅改配置，不改调用方。
+    """
+
+    #: provider 名称
+    name: str = "base"
+    #: 是否走公网（出域）。prod 环境禁止 egress=True 的 provider。
+    egress: bool = False
+
+    @abc.abstractmethod
+    def chat(self, messages: list[ChatMessage], **kwargs) -> LLMResponse:
+        """同步对话补全。"""
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def health(self) -> bool:
+        """探活：provider 是否可用。"""
+        raise NotImplementedError

backend/app/llm/factory.py

@@ -0,0 +1,31 @@
+"""LLM Provider 工厂：按配置创建 provider，并执行数据零出域红线校验。"""
+
+from __future__ import annotations
+
+from app.config import EGRESS_PROVIDERS, LLMProviderName, Settings, get_settings
+from app.llm.base import LLMProvider
+from app.llm.providers import DashScopeProvider, VllmProvider
+
+
+class EgressPolicyError(RuntimeError):
+    """数据零出域红线违规。"""
+
+
+def get_llm_provider(settings: Settings | None = None) -> LLMProvider:
+    settings = settings or get_settings()
+
+    # 红线：prod 环境禁止公网 provider
+    if settings.is_prod and settings.llm_provider in EGRESS_PROVIDERS:
+        raise EgressPolicyError(
+            f"数据零出域红线违规：prod 环境禁止使用公网 LLM Provider "
+            f"'{settings.llm_provider.value}'。"
+        )
+
+    if settings.llm_provider == LLMProviderName.dashscope:
+        return DashScopeProvider(
+            api_key=settings.dashscope_api_key, model=settings.dashscope_model
+        )
+    if settings.llm_provider == LLMProviderName.vllm:
+        return VllmProvider(base_url=settings.vllm_base_url, model=settings.vllm_model)
+
+    raise ValueError(f"未知的 LLM Provider: {settings.llm_provider}")

backend/app/llm/providers.py

@@ -0,0 +1,80 @@
+"""具体 LLM Provider 实现：DashScope（公网千问，仅 dev）、vLLM（本地，prod）。
+
+两者均走 OpenAI 兼容的 /chat/completions 协议。
+"""
+
+from __future__ import annotations
+
+import httpx
+
+from app.llm.base import ChatMessage, LLMProvider, LLMResponse
+
+
+class DashScopeProvider(LLMProvider):
+    """公网千问（DashScope，OpenAI 兼容模式）。仅限开发测试，且只允许脱敏/样例假数据。"""
+
+    name = "dashscope"
+    egress = True  # 走公网，出域
+
+    _BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+
+    def __init__(self, api_key: str, model: str, timeout: float = 30.0) -> None:
+        self._api_key = api_key
+        self._model = model
+        self._timeout = timeout
+
+    def chat(self, messages: list[ChatMessage], **kwargs) -> LLMResponse:
+        payload = {
+            "model": self._model,
+            "messages": [{"role": m.role, "content": m.content} for m in messages],
+            **kwargs,
+        }
+        headers = {"Authorization": f"Bearer {self._api_key}"}
+        with httpx.Client(timeout=self._timeout) as client:
+            resp = client.post(
+                f"{self._BASE_URL}/chat/completions", json=payload, headers=headers
+            )
+            resp.raise_for_status()
+            data = resp.json()
+        content = data["choices"][0]["message"]["content"]
+        return LLMResponse(
+            content=content, model=self._model, provider=self.name, egress=True, raw=data
+        )
+
+    def health(self) -> bool:
+        return bool(self._api_key)
+
+
+class VllmProvider(LLMProvider):
+    """本地 vLLM（OpenAI 兼容）。生产使用，数据不出域。"""
+
+    name = "vllm"
+    egress = False
+
+    def __init__(self, base_url: str, model: str, timeout: float = 60.0) -> None:
+        self._base_url = base_url.rstrip("/")
+        self._model = model
+        self._timeout = timeout
+
+    def chat(self, messages: list[ChatMessage], **kwargs) -> LLMResponse:
+        payload = {
+            "model": self._model,
+            "messages": [{"role": m.role, "content": m.content} for m in messages],
+            **kwargs,
+        }
+        with httpx.Client(timeout=self._timeout) as client:
+            resp = client.post(f"{self._base_url}/chat/completions", json=payload)
+            resp.raise_for_status()
+            data = resp.json()
+        content = data["choices"][0]["message"]["content"]
+        return LLMResponse(
+            content=content, model=self._model, provider=self.name, egress=False, raw=data
+        )
+
+    def health(self) -> bool:
+        try:
+            with httpx.Client(timeout=5.0) as client:
+                resp = client.get(f"{self._base_url}/models")
+                return resp.status_code == 200
+        except httpx.HTTPError:
+            return False