完善风险评分、LLM兜底与会话历史加载

- 接入数据库驱动的七维风险评分第一阶段逻辑\n- 增加主模型失败后的 DeepSeek 语义、流式和预防建议兜底\n- 将 AI 预防建议改为基于评分、历史异常和标准规则动态生成\n- 增加无历史案例时的标准库兜底校验和通用规则匹配\n- 修复未知供应商识别及测试占位名称过滤\n- 增加后端滚动日志文件配置和 LLM 请求日志\n- 历史会话直接读取已保存结果，避免重新调用模型和重算风险\n- 补充评分、供应商识别和 LLM 协议测试

.env.example

@@ -1,19 +1,52 @@
+# ==================== 应用基础配置 ====================
+# 应用名称、版本和运行环境，用于接口文档、日志和运行行为识别。
 RISK_APP_NAME=Dieya Risk Prediction Backend
 RISK_APP_VERSION=0.1.0
 RISK_ENVIRONMENT=development
 RISK_DEBUG=false
+
+# ==================== 数据库配置 ====================
+# 风险系统业务数据库连接地址，包含数据库用户、密码、主机、端口和库名。
 RISK_DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5432/dieya_risk_prediction
+
+# ==================== 日志配置 ====================
+# 控制后端日志级别；可选 DEBUG、INFO、WARNING、ERROR。
 RISK_LOG_LEVEL=INFO
+
+# ==================== JWT 登录鉴权配置 ====================
+# JWT 签名密钥、签名算法和登录态有效期（分钟）。生产环境必须更换签名密钥。
 RISK_JWT_SECRET_KEY=CHANGE_THIS_IN_PRODUCTION
 RISK_JWT_ALGORITHM=HS256
 RISK_JWT_EXPIRE_MINUTES=480
+
+# ==================== 钉钉免登配置 ====================
+# 用于钉钉免登和组织身份校验；未接入钉钉时可以留空。
 RISK_DINGTALK_APP_KEY=
 RISK_DINGTALK_APP_SECRET=
 RISK_DINGTALK_CORP_ID=
 RISK_DINGTALK_AGENT_ID=
+
+# ==================== 主 LLM 模型配置 ====================
+# 主模型 API 密钥、OpenAI 兼容接口地址、模型名、请求超时时间和生成温度。
 RISK_LLM_API_KEY=
 RISK_LLM_BASE_URL=https://api.openai.com/v1
 RISK_LLM_MODEL=gpt-4o-mini
 RISK_LLM_TIMEOUT_SECONDS=120
 RISK_LLM_TEMPERATURE=0.2
+
+# ==================== LLM 兜底模型配置 ====================
+# 主模型请求失败时使用 DeepSeek；填写独立的 DeepSeek API Key 才会真正启用兜底。
+# API_MODE=responses 时调用 /responses；如服务商只支持 Chat Completions，可改为 chat_completions。
+# MAX_RETRIES 是兜底端点内部的重试次数，REASONING_EFFORT 适用于 Responses 推理配置。
+RISK_LLM_FALLBACK_API_KEY=
+RISK_LLM_FALLBACK_BASE_URL=https://api.deepseek.com
+RISK_LLM_FALLBACK_MODEL=deepseek-v4-pro
+RISK_LLM_FALLBACK_API_MODE=responses
+RISK_LLM_FALLBACK_TIMEOUT_SECONDS=30
+RISK_LLM_FALLBACK_MAX_RETRIES=1
+RISK_LLM_FALLBACK_REASONING_EFFORT=high
+RISK_LLM_FALLBACK_MAX_OUTPUT_TOKENS=16384
+
+# ==================== 跨域配置 ====================
+# 允许访问后端接口的前端地址，多个地址用英文逗号分隔。
 RISK_CORS_ALLOW_ORIGINS=http://localhost:39158,http://127.0.0.1:39158

.gitignore

@@ -6,8 +6,8 @@ __pycache__/
 .coverage
 htmlcov/
 .env
+logs/
 *.egg-info/
 dist/
 build/
 data/
-

app/api/v1/risk_predictions.py

@@ -4,7 +4,7 @@ from typing import Annotated
 
 from fastapi import APIRouter, Depends
 
-from app.api.dependencies import require_permission
+from app.api.dependencies import DatabaseSession, require_permission
 from app.models.entities import UserAccount
 from app.schemas.common import ApiResponse
 from app.schemas.risk_prediction import ModelInfo, RiskAssessmentRequest, RiskAssessmentResult
@@ -23,6 +23,7 @@ async def model_info(
 @router.post("/preview", response_model=ApiResponse[RiskAssessmentResult])
 async def preview_risk(
     request: RiskAssessmentRequest,
+    session: DatabaseSession,
     _: Annotated[UserAccount, Depends(require_permission("risk.assess", operate=True))],
 ) -> ApiResponse[RiskAssessmentResult]:
-    return ApiResponse(data=RiskPredictionService.preview(request))
+    return ApiResponse(data=await RiskPredictionService.preview(request, session))

app/core/config.py

@@ -1,7 +1,7 @@
 """Environment-backed application settings."""
 
 from functools import lru_cache
-from typing import Annotated
+from typing import Annotated, Literal
 
 from pydantic import AliasChoices, Field, SecretStr, field_validator
 from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict
@@ -53,6 +53,45 @@ class Settings(BaseSettings):
         le=2,
         validation_alias=AliasChoices("RISK_LLM_TEMPERATURE", "LLM_TEMPERATURE"),
     )
+    llm_fallback_api_key: SecretStr | None = Field(
+        default=None,
+        repr=False,
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_API_KEY", "DEEPSEEK_API_KEY"),
+    )
+    llm_fallback_base_url: str = Field(
+        default="https://api.deepseek.com",
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_BASE_URL", "DEEPSEEK_BASE_URL"),
+    )
+    llm_fallback_model: str = Field(
+        default="deepseek-v4-pro",
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_MODEL", "DEEPSEEK_MODEL"),
+    )
+    llm_fallback_api_mode: Literal["chat_completions", "responses"] = Field(
+        default="responses",
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_API_MODE", "DEEPSEEK_API_MODE"),
+    )
+    llm_fallback_timeout_seconds: int = Field(
+        default=30,
+        gt=0,
+        le=600,
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_TIMEOUT_SECONDS", "DEEPSEEK_TIMEOUT"),
+    )
+    llm_fallback_max_retries: int = Field(
+        default=1,
+        ge=0,
+        le=10,
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_MAX_RETRIES", "DEEPSEEK_MAX_RETRIES"),
+    )
+    llm_fallback_reasoning_effort: Literal["low", "high", "max"] = Field(
+        default="high",
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_REASONING_EFFORT", "DEEPSEEK_REASONING_EFFORT"),
+    )
+    llm_fallback_max_output_tokens: int = Field(
+        default=16_384,
+        ge=1,
+        le=393_216,
+        validation_alias=AliasChoices("RISK_LLM_FALLBACK_MAX_OUTPUT_TOKENS", "DEEPSEEK_MAX_OUTPUT_TOKENS"),
+    )
     cors_allow_origins: Annotated[tuple[str, ...], NoDecode] = (
         "http://localhost:39158",
         "http://127.0.0.1:39158",
@@ -77,6 +116,11 @@ class Settings(BaseSettings):
     def normalize_llm_base_url(cls, value: str) -> str:
         return value.rstrip("/")
 
+    @field_validator("llm_fallback_base_url")
+    @classmethod
+    def normalize_llm_fallback_base_url(cls, value: str) -> str:
+        return value.rstrip("/")
+
     @field_validator("cors_allow_origins", mode="before")
     @classmethod
     def split_cors_origins(cls, value: object) -> object:

app/core/logging.py

@@ -1,20 +1,60 @@
 """Shared logging setup."""
 
 import logging
+from logging.handlers import RotatingFileHandler
+from pathlib import Path
 
 LOGGER_NAME = "dieya_risk_prediction"
+LOG_DIR = Path(__file__).resolve().parents[2] / "logs"
+LOG_FILE = LOG_DIR / "backend.log"
+_FORMAT = "%(asctime)s %(levelname)s %(name)s %(message)s"
 
 
 def configure_logging(level: str = "INFO") -> logging.Logger:
     logger = logging.getLogger(LOGGER_NAME)
     logger.setLevel(level.upper())
     logger.propagate = False
-    if not logger.handlers:
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
-        handler = logging.StreamHandler()
+    formatter = logging.Formatter(_FORMAT)
-        handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)s %(name)s %(message)s"))
+    file_handler = next(
-        logger.addHandler(handler)
+        (
+            handler
+            for handler in logger.handlers
+            if getattr(handler, "_dieya_file_handler", False)
+        ),
+        None,
+    )
+    if file_handler is None:
+        file_handler = RotatingFileHandler(
+            LOG_FILE,
+            maxBytes=10 * 1024 * 1024,
+            backupCount=5,
+            encoding="utf-8",
+        )
+        file_handler._dieya_file_handler = True
+        file_handler.setFormatter(formatter)
+        logger.addHandler(file_handler)
+    root_logger = logging.getLogger()
+    root_logger.setLevel(level.upper())
+    if not any(getattr(handler, "_dieya_file_handler", False) for handler in root_logger.handlers):
+        root_file_handler = RotatingFileHandler(
+            LOG_FILE,
+            maxBytes=10 * 1024 * 1024,
+            backupCount=5,
+            encoding="utf-8",
+        )
+        root_file_handler._dieya_file_handler = True
+        root_file_handler.setFormatter(formatter)
+        root_logger.addHandler(root_file_handler)
+    if not any(
+        isinstance(handler, logging.StreamHandler)
+        and not getattr(handler, "_dieya_file_handler", False)
+        for handler in logger.handlers
+    ):
+        stream_handler = logging.StreamHandler()
+        stream_handler.setFormatter(formatter)
+        logger.addHandler(stream_handler)
     return logger
 
 
 logger = configure_logging()
-

app/repositories/conversations.py

@@ -44,7 +44,6 @@ class ConversationRepository:
                 Conversation.id == conversation_id,
                 Conversation.user_id == user_id,
             )
-            .options(selectinload(Conversation.messages))
         )
         if conversation is None:
             return None

app/schemas/risk_prediction.py

@@ -14,7 +14,13 @@ class RiskIndicator(ApiSchema):
 
 class RiskAssessmentRequest(ApiSchema):
     subject_id: str = Field(min_length=1, max_length=128)
-    indicators: list[RiskIndicator] = Field(min_length=1, max_length=50)
+    indicators: list[RiskIndicator] = Field(default_factory=list, max_length=50)
+    product_category: str | None = Field(default=None, max_length=100)
+    material_type: str | None = Field(default=None, max_length=200)
+    supplier: str | None = Field(default=None, max_length=300)
+    factory: str | None = Field(default=None, max_length=300)
+    stage_new: str | None = Field(default=None, max_length=100)
+    defect_level: str | None = Field(default=None, max_length=10)
 
 
 class RiskFactorResult(ApiSchema):
@@ -54,6 +60,10 @@ class RiskDimensionScore(ApiSchema):
     dimension: str
     score: float
     weight: float
+    status: str = "database"
+    reason: str | None = None
+    sample_size: int | None = None
+    raw_value: float | int | None = None
 
 
 class SimilarCase(ApiSchema):

app/services/conversations.py

@@ -1,6 +1,7 @@
 """Conversation use cases and the initial risk-assistant response adapter."""
 
 import json
+import logging
 import re
 from collections.abc import AsyncGenerator, Sequence
 from datetime import UTC, datetime
@@ -25,8 +26,11 @@ from app.schemas.risk_prediction import (
     StandardJudgment,
     SupplierFrequency,
 )
+from app.services.database_risk_scoring import DatabaseRiskScoringService
 from app.services.llm_client import LLMClient
 
+logger = logging.getLogger("dieya_risk_prediction.conversations")
+
 
 class ConversationService:
     _engine = BaselineRiskEngine()
@@ -93,13 +97,11 @@ class ConversationService:
         if page is None:
             raise AppException("对话不存在或无权访问", code=404, status_code=404)
         conversation, messages, total = page
-        live_results = await self._refresh_risk_results(messages)
         return self._data(
             conversation,
             messages=messages,
             total=total,
             offset=offset,
-            risk_results=live_results,
         )
 
     async def create(self, user: UserAccount, title: str | None = None) -> ConversationData:
@@ -160,13 +162,11 @@ class ConversationService:
         if page is None:  # pragma: no cover - the row already existed
             raise AppException("对话读取失败", code=500, status_code=500)
         persisted, messages, total = page
-        live_results = await self._refresh_risk_results(messages)
         return self._data(
             persisted,
             messages=messages,
             total=total,
             offset=0,
-            risk_results=live_results,
         )
 
     async def stream(
@@ -208,6 +208,11 @@ class ConversationService:
             yield {"type": "error", "message": str(exc)}
             return
         semantics = self._normalize_semantics(text, semantics)
+        semantics = await self._normalize_follow_up_semantics(
+            text,
+            semantics,
+            conversation,
+        )
         yield {"type": "semantic", "data": semantics}
 
         context = self._llm_messages(conversation, text, semantics)
@@ -228,6 +233,11 @@ class ConversationService:
 
         show_risk = self._is_risk_request(text, semantics)
         response_type = self._response_type(text, semantics)
+        response_type = await self._use_solution_card_without_history(
+            response_type,
+            text,
+            semantics,
+        )
         risk_result = (
             await self._enrich_risk_result(text, semantics, self._predict(text))
             if show_risk and response_type == "risk_assessment"
@@ -265,13 +275,11 @@ class ConversationService:
             yield {"type": "error", "message": "对话读取失败"}
             return
         persisted, messages, total = page
-        live_results = await self._refresh_risk_results(messages)
         result = self._data(
             persisted,
             messages=messages,
             total=total,
             offset=0,
-            risk_results=live_results,
         )
         yield {"type": "done", "data": result.model_dump(mode="json", by_alias=True)}
 
@@ -309,6 +317,8 @@ class ConversationService:
     def _response_type(cls, text: str, semantics: dict[str, Any]) -> str:
         """Choose the prototype card from the user's actual question."""
 
+        if cls._is_placeholder_supplier_name(text):
+            return "general"
         intent = str(semantics.get("intent") or "").strip().lower()
         if intent == "supplier" or (
             any(word in text for word in ("供应商", "加工厂", "工厂"))
@@ -322,15 +332,138 @@ class ConversationService:
             return "standard"
         if intent == "compliance" or any(word in text for word in ("合规", "宣称", "备案", "法规")):
             return "compliance"
-        if intent == "solution" or (
+        if intent == "solution":
-            any(word in text for word in ("次氯酸", "消毒喷雾"))
-            and not any(word in text for word in ("评估", "风险评估", "风险预估"))
-        ):
             return "solution"
         if intent == "assessment" or cls._is_explicit_risk_request(text):
             return "risk_assessment"
         return "general"
 
+    async def _normalize_follow_up_semantics(
+        self,
+        text: str,
+        semantics: dict[str, Any],
+        conversation: Conversation,
+    ) -> dict[str, Any]:
+        """Recover card intent for short follow-ups such as ``佳亿``.
+
+        The LLM can answer a short follow-up using conversation history while
+        still labeling the current turn as ``general``. Card routing needs a
+        deterministic signal, so known supplier names are matched locally and
+        written back into the semantic payload before routing.
+        """
+
+        normalized = dict(semantics)
+        if self._is_placeholder_supplier_name(text):
+            normalized.update(
+                {
+                    "intent": "general",
+                    "risk_requested": False,
+                    "semantic_source": "placeholder_filtered",
+                }
+            )
+            normalized.pop("supplier", None)
+            return normalized
+        intent = str(normalized.get("intent") or "").strip().lower()
+        if intent not in {"", "general", "greet"}:
+            return normalized
+        supplier_names = list(
+            (
+                await self.session.scalars(
+                    select(AnomalyCase.supplier)
+                    .where(AnomalyCase.supplier.is_not(None))
+                    .distinct()
+                )
+            ).all()
+        )
+        supplier_name = self._infer_supplier_follow_up(text, supplier_names)
+        supplier_name = supplier_name or self._infer_unknown_supplier_follow_up(text)
+        if not supplier_name:
+            return normalized
+        normalized.update(
+            {
+                "intent": "supplier",
+                "risk_requested": True,
+                "supplier": supplier_name,
+                "confidence": max(float(normalized.get("confidence") or 0), 0.75),
+                "semantic_source": "database_supplier_match"
+                if supplier_name in supplier_names
+                else "standalone_new_supplier_fallback",
+            }
+        )
+        return normalized
+
+    @staticmethod
+    def _infer_supplier_follow_up(text: str, supplier_names: Sequence[str | None]) -> str | None:
+        """Return a known supplier when the current message names it directly."""
+
+        cleaned = text.strip().lower()
+        if not cleaned:
+            return None
+        candidates = sorted(
+            {
+                " ".join(name.strip().split())
+                for name in supplier_names
+                if name and name.strip() and name.strip() not in {"/", "-"}
+            },
+            key=len,
+            reverse=True,
+        )
+        for name in candidates:
+            if cleaned == name.lower():
+                return name
+        return None
+
+    @staticmethod
+    def _infer_unknown_supplier_follow_up(text: str) -> str | None:
+        """Treat a short standalone business name as a new supplier.
+
+        This is deliberately conservative: common greetings, product terms,
+        workflow commands, and messages containing punctuation are excluded.
+        Unknown names are shown as a new supplier with zero historical data.
+        """
+
+        cleaned = " ".join(text.strip().split())
+        if not 2 <= len(cleaned) <= 20 or " " in cleaned:
+            return None
+        if not re.fullmatch(r"[\u4e00-\u9fffA-Za-z][\u4e00-\u9fffA-Za-z0-9_-]*", cleaned):
+            return None
+        excluded = {
+            "你好",
+            "您好",
+            "谢谢",
+            "好的",
+            "继续",
+            "继续分析",
+            "洗发水",
+            "沐浴露",
+            "面霜",
+            "脱毛膏",
+            "清洁剂",
+            "牙膏",
+            "软管",
+            "泵头",
+            "瓶子",
+            "彩盒",
+            "标准库",
+            "风险评估",
+            "供应商",
+            "新供应商A",
+            "测试供应商",
+        }
+        if (
+            cleaned in excluded
+            or cleaned.startswith("测试")
+            or cleaned.startswith("新供应商")
+            or any(term in cleaned for term in ("请", "帮我", "分析", "查询", "怎么样"))
+        ):
+            return None
+        return cleaned
+
+    @staticmethod
+    def _is_placeholder_supplier_name(text: str) -> bool:
+        cleaned = " ".join(text.strip().split())
+        return cleaned.startswith("测试") or cleaned.startswith("新供应商")
+
     async def _build_response_card(
         self,
         response_type: str,
@@ -347,6 +480,59 @@ class ConversationService:
         builder = builders.get(response_type)
         return await builder(text, semantics) if builder else None
 
+    async def _use_solution_card_without_history(
+        self,
+        response_type: str,
+        text: str,
+        semantics: dict[str, Any],
+    ) -> str:
+        """Use the solution card for a product risk question with no history."""
+
+        if response_type != "risk_assessment":
+            return response_type
+        category = self._semantic_text(semantics, "category") or self._first_match(
+            text,
+            (
+                "洗发水",
+                "沐浴露",
+                "面霜",
+                "乳液",
+                "精华",
+                "面膜",
+                "防晒",
+                "护手霜",
+                "护发素",
+                "脱毛膏",
+                "清洁剂",
+                "牙膏",
+                "消毒喷雾",
+                "次氯酸",
+            ),
+        )
+        has_product_context = bool(
+            category
+            or any(
+                key in text
+                for key in (
+                    "新品",
+                    "新产品",
+                    "产品开发",
+                    "产品立项",
+                    "规格",
+                    "容量",
+                    "包装",
+                    "材质",
+                    "配方",
+                    "原料",
+                )
+            )
+        )
+        if not has_product_context:
+            return response_type
+        if not category or await self._count_category_anomalies(category) == 0:
+            return "solution"
+        return response_type
+
     async def _supplier_card(self, text: str, semantics: dict[str, Any]) -> dict[str, Any]:
         name = self._semantic_text(semantics, "supplier") or self._semantic_text(semantics, "factory")
         if not name:
@@ -451,10 +637,29 @@ class ConversationService:
 
     async def _solution_card(self, text: str, semantics: dict[str, Any]) -> dict[str, Any]:
         category = self._semantic_text(semantics, "category") or self._first_match(
-            text, ("次氯酸消毒喷雾", "消毒喷雾", "清洁剂", "脱毛膏")
+            text,
+            (
+                "次氯酸消毒喷雾",
+                "消毒喷雾",
+                "清洁剂",
+                "脱毛膏",
+                "洗发水",
+                "沐浴露",
+                "面霜",
+                "乳液",
+                "精华",
+                "面膜",
+                "防晒",
+                "护手霜",
+                "护发素",
+                "牙膏",
+            ),
         ) or "当前新品"
         direct_count = await self._count_category_anomalies(category)
         standards = await self._standards_for(category, text)
+        standard_fallback_enabled = direct_count == 0
+        if standard_fallback_enabled and not standards:
+            standards = await self._fallback_standards_for(text)
         risks: list[dict[str, str]] = []
         if "次氯酸" in text or "铝" in text:
             risks.append(
@@ -480,6 +685,31 @@ class ConversationService:
                     "detail": "供应商或加工厂的历史表现需要结合异常案例和现场审核确认。",
                 }
             )
+        fallback_suggestions = [
+            "小样阶段完成内容物与包装材料相容性测试",
+            "增加倒置密封测试和加速老化测试",
+            "首单安排来料全检，并根据结果决定是否驻厂巡检",
+            "实际执行前由品控确认并安排小批量试产验证",
+        ]
+        try:
+            suggestions = await self.llm.generate_prevention_suggestions(
+                {
+                    "user_input": text,
+                    "category": category,
+                    "historical_case_count": direct_count,
+                    "standard_fallback_enabled": standard_fallback_enabled,
+                    "matched_standards": [self._standard_dict(item) for item in standards[:5]],
+                    "identified_risks": risks,
+                    "rag_similar_cases": [],
+                }
+            )
+        except Exception as exc:  # noqa: BLE001 - the solution card must remain available
+            logger.warning(
+                "Solution card AI suggestions failed error=%s; using deterministic fallback",
+                exc.__class__.__name__,
+            )
+            suggestions = fallback_suggestions
+
         return {
             "type": "solution",
             "title": "AI 案例推理生成",
@@ -488,19 +718,76 @@ class ConversationService:
                 f"目标品类「{category}」当前命中历史异常 {direct_count} 条；"
                 "方案结合已导入异常案例和标准规则生成。"
                 if direct_count
-                else f"目标品类「{category}」暂无直接历史案例，当前按标准规则库进行兜底分析。"
+                else f"目标品类「{category}」暂无直接历史案例，当前按技术法规规则库进行兜底分析。"
             ),
+            "standardFallbackEnabled": standard_fallback_enabled,
             "standards": [self._standard_dict(item) for item in standards[:4]],
             "risks": risks,
-            "suggestions": [
+            "suggestions": suggestions,
-                "小样阶段完成内容物与包装材料相容性测试",
-                "增加倒置密封测试和加速老化测试",
-                "首单安排来料全检，并根据结果决定是否驻厂巡检",
-                "实际执行前由品控确认并安排小批量试产验证",
-            ],
             "warning": "以上方案由 AI 推理生成，实际执行前请咨询品控部门确认。",
         }
 
+    async def _fallback_standards_for(self, text: str) -> Sequence[StandardJudgment]:
+        """Find generic safety rules when no direct product history is available."""
+
+        terms = [
+            term
+            for term in (
+                self._first_match(text, ("相容性", "密封", "稳定性", "消毒", "微生物", "氧化")),
+                self._first_match(text, ("铝", "金属", "喷头", "软管", "PET", "泵头")),
+            )
+            if term
+        ]
+        if terms:
+            clauses = [
+                or_(
+                    StandardRule.category_lv1.ilike(f"%{term}%"),
+                    StandardRule.category_lv2.ilike(f"%{term}%"),
+                    StandardRule.category_lv3.ilike(f"%{term}%"),
+                    StandardRule.risk_point.ilike(f"%{term}%"),
+                    StandardRule.judge_standard.ilike(f"%{term}%"),
+                )
+                for term in terms
+            ]
+            rows = list(
+                await self.session.scalars(
+                    select(StandardRule)
+                    .where(StandardRule.status == "启用", or_(*clauses))
+                    .order_by(StandardRule.id)
+                    .limit(4)
+                )
+            )
+            if rows:
+                return [
+                    StandardJudgment(
+                        category=row.category_lv1,
+                        risk_point=row.risk_point,
+                        rule=row.judge_standard,
+                        source=row.regulation or "标准文件.xlsx",
+                        status="待检测",
+                    )
+                    for row in rows
+                ]
+
+        rows = list(
+            await self.session.scalars(
+                select(StandardRule)
+                .where(StandardRule.status == "启用")
+                .order_by(StandardRule.id)
+                .limit(4)
+            )
+        )
+        return [
+            StandardJudgment(
+                category=row.category_lv1,
+                risk_point=row.risk_point,
+                rule=row.judge_standard,
+                source=row.regulation or "标准文件.xlsx",
+                status="待检测",
+            )
+            for row in rows
+        ]
+
     async def _top_anomaly_types(self, column, value: str | None) -> Sequence[dict[str, Any]]:  # noqa: ANN001
         if not value:
             return []
@@ -603,20 +890,29 @@ class ConversationService:
             supplier = None
         supplier_or_factory = supplier or factory
 
+        database_score = await DatabaseRiskScoringService(self.session).score(
+            category=category,
+            material=self._first_match(
+                text,
+                ("软管", "PET瓶", "PET", "铝瓶", "泵头", "彩盒", "瓶子", "喷头", "盖子"),
+            ),
+            supplier=supplier,
+            factory=factory,
+            stage=self._semantic_text(semantics, "stage") or self._first_match(
+                text, ("需求发现", "产品立项", "产品开发", "量产上市", "产品生命管理")
+            ),
+            defect_level=self._first_match(text, ("A级", "B级", "C级")),
+        )
         category_count = await self._count_category_anomalies(category)
-        supplier_count = await self._count_anomalies(AnomalyCase.supplier, supplier_or_factory)
+        supplier_count = await self._count_anomalies(AnomalyCase.supplier, supplier)
         factory_count = await self._count_anomalies(AnomalyCase.factory, factory)
-        if factory and not supplier:
+        anomaly_count = supplier_count if supplier else factory_count
-            anomaly_count = factory_count
-        elif supplier and not factory:
-            anomaly_count = supplier_count
-        else:
-            anomaly_count = supplier_count or factory_count
         standard_rows = await self._standards_for(category, text)
 
-        alpha = round(min(1.5, 1 + (anomaly_count / 100) * 0.3), 2) if anomaly_count else 1.0
+        alpha_count = supplier_count if supplier else 0
-        tier = self._supplier_tier(anomaly_count)
+        alpha = round(min(1.5, 1 + (alpha_count / 100) * 0.3), 2) if alpha_count else 1.0
-        level_text = self._risk_level_text(result.level)
+        tier = self._supplier_tier(alpha_count) if supplier else None
+        level_text = self._risk_level_text(database_score.level)
         alert_text = self._alert_text(
             category,
             category_count,
@@ -626,17 +922,45 @@ class ConversationService:
         )
         dimensions = [
             RiskDimensionScore(
-                dimension=self._dimension_name(factor.name),
+                dimension=item.name,
-                score=round(max(0, 100 - factor.value), 2),
+                score=item.score,
-                weight=factor.weight,
+                weight=item.weight,
+                status=item.status,
+                reason=item.reason,
+                sample_size=item.sample_size,
+                raw_value=item.raw_value,
             )
-            for factor in result.factors
+            for item in database_score.dimensions
         ]
-        suggestions = self._suggestions(text, supplier_or_factory)
+        factors = [
+            RiskFactorResult(
+                name=item.name,
+                value=round(100 - item.score, 2),
+                weight=item.weight,
+                contribution=round(item.score * item.weight / 100, 2),
+            )
+            for item in database_score.dimensions
+        ]
+        suggestions = await self._generate_ai_suggestions(
+            text=text,
+            semantics=semantics,
+            category=category,
+            supplier_or_factory=supplier_or_factory,
+            category_count=category_count,
+            anomaly_count=anomaly_count,
+            alpha=alpha,
+            database_score=database_score,
+            dimensions=dimensions,
+            standard_rows=standard_rows,
+        )
 
         return result.model_copy(
             update={
-                "risk_badge": self._risk_badge(result.level),
+                "score": database_score.score,
+                "level": database_score.level,
+                "model_version": "database-seven-dimension-v1",
+                "factors": factors,
+                "risk_badge": self._risk_badge(database_score.level),
                 "alert_text": alert_text,
                 "category_warning": CategoryWarning(
                     category=category,
@@ -675,6 +999,62 @@ class ConversationService:
         )
         return int(count or 0)
 
+    async def _generate_ai_suggestions(
+        self,
+        *,
+        text: str,
+        semantics: dict[str, Any],
+        category: str,
+        supplier_or_factory: str | None,
+        category_count: int,
+        anomaly_count: int,
+        alpha: float,
+        database_score,
+        dimensions: Sequence[RiskDimensionScore],
+        standard_rows: Sequence[StandardJudgment],
+    ) -> Sequence[str]:  # noqa: ANN001
+        """Generate suggestions from live scoring data, with a deterministic fallback."""
+
+        context = {
+            "user_input": text,
+            "semantic_fields": {
+                key: value
+                for key, value in semantics.items()
+                if key in {"category", "specification", "packaging", "supplier", "factory", "stage"}
+                and value
+            },
+            "category": category,
+            "supplier_or_factory": supplier_or_factory,
+            "category_anomaly_count": category_count,
+            "supplier_or_factory_anomaly_count": anomaly_count,
+            "frequency_alpha": alpha,
+            "overall_score": database_score.score,
+            "risk_level": self._risk_level_text(database_score.level),
+            "dimensions": [
+                {
+                    "name": item.dimension,
+                    "score": item.score,
+                    "weight": item.weight,
+                    "status": item.status,
+                    "reason": item.reason,
+                    "sample_size": item.sample_size,
+                }
+                for item in dimensions
+            ],
+            "matched_standards": [self._standard_dict(item) for item in standard_rows[:5]],
+            "rag_similar_cases": [],
+        }
+        try:
+            suggestions = await self.llm.generate_prevention_suggestions(context)
+            logger.info("AI prevention suggestions generated count=%s", len(suggestions))
+            return suggestions
+        except Exception as exc:  # noqa: BLE001 - suggestions must not break risk assessment
+            logger.warning(
+                "AI prevention suggestion generation failed error=%s; using deterministic fallback",
+                exc.__class__.__name__,
+            )
+            return self._suggestions(text, supplier_or_factory)
+
     async def _count_category_anomalies(self, category: str) -> int:
         """The imported workbook has sparse product_category values.
 
@@ -878,9 +1258,7 @@ class ConversationService:
             id=conversation.id,
             title=conversation.title,
             message_count=resolved_total,
-            last_risk_level=ConversationService._last_risk_level_from_messages(
+            last_risk_level=ConversationService._last_risk_level_from_messages(resolved_messages),
-                resolved_messages,
-            ) or ConversationService._last_risk_level(conversation),
             updated_at=conversation.updated_at,
             messages=[
                 cls._message(
@@ -898,41 +1276,6 @@ class ConversationService:
             has_more_messages=offset + len(resolved_messages) < resolved_total,
         )
 
-    async def _refresh_risk_results(
-        self,
-        messages: Sequence[ConversationMessage],
-    ) -> dict[int, RiskAssessmentResult]:
-        """Refresh database-backed card fields for historical messages.
-
-        Risk scores are intentionally kept as the persisted assessment result,
-        while fields sourced from imported cases and standards are re-read from
-        PostgreSQL. This prevents an old conversation from showing stale zero
-        counts after the reference workbooks have been imported.
-        """
-
-        refreshed: dict[int, RiskAssessmentResult] = {}
-        for index, message in enumerate(messages):
-            if message.role != "assistant":
-                continue
-            result = self._risk_result_for_message(messages, index)
-            if result is None:
-                continue
-            user_text = next(
-                (
-                    candidate.content
-                    for candidate in reversed(messages[:index])
-                    if candidate.role == "user"
-                ),
-                result.subject_id,
-            )
-            semantic = (message.structured_data or {}).get("semantic")
-            refreshed[message.id] = await self._enrich_risk_result(
-                user_text,
-                semantic if isinstance(semantic, dict) else {},
-                result,
-            )
-        return refreshed
-
     @staticmethod
     def _message(
         message: ConversationMessage,

app/services/database_risk_scoring.py

@@ -0,0 +1,406 @@
+"""First-phase database-backed seven-dimension risk scoring."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import date, timedelta
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.domain.risk_engine import RiskLevel
+from app.models.entities import AnomalyCase
+
+DEFAULT_WEIGHTS = {
+    "供应商表现": 25.0,
+    "物料品类": 20.0,
+    "产品品类": 15.0,
+    "加工厂质量": 15.0,
+    "标准符合性": 10.0,
+    "节点阶段": 10.0,
+    "缺陷等级": 5.0,
+}
+
+STAGES = ("需求发现", "产品立项", "产品开发", "量产上市", "产品生命管理")
+
+
+@dataclass(frozen=True, slots=True)
+class DimensionScore:
+    name: str
+    score: float
+    weight: float
+    status: str
+    reason: str | None = None
+    sample_size: int | None = None
+    raw_value: float | int | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class DatabaseRiskScore:
+    score: float
+    level: RiskLevel
+    dimensions: tuple[DimensionScore, ...]
+
+
+class DatabaseRiskScoringService:
+    """Calculate the PRD first-phase model from anomaly_cases.
+
+    Inspection results and standard test results are intentionally not guessed
+    here. Their dimensions use the PRD neutral score of 50 and expose the
+    reason so callers can distinguish a real score from a fallback.
+    """
+
+    def __init__(self, session: AsyncSession, *, as_of: date | None = None) -> None:
+        self.session = session
+        self.as_of = as_of or date.today()
+
+    async def score(
+        self,
+        *,
+        category: str | None = None,
+        material: str | None = None,
+        supplier: str | None = None,
+        factory: str | None = None,
+        stage: str | None = None,
+        defect_level: str | None = None,
+    ) -> DatabaseRiskScore:
+        rows = await self._load_rows()
+        category_name = normalize_category(category)
+        material_name = normalize_material(material)
+        stage_name = normalize_stage(stage)
+        party_name = clean_value(supplier)
+
+        dimensions = (
+            self._supplier_dimension(rows, party_name),
+            self._material_dimension(rows, material_name),
+            self._category_dimension(rows, category_name),
+            DimensionScore(
+                name="加工厂质量",
+                score=50.0,
+                weight=DEFAULT_WEIGHTS["加工厂质量"],
+                status="default",
+                reason="当前数据库没有巡检合格率/批次数据，按 PRD 中性值 50 分计入",
+            ),
+            DimensionScore(
+                name="标准符合性",
+                score=50.0,
+                weight=DEFAULT_WEIGHTS["标准符合性"],
+                status="default",
+                reason="当前数据库没有检测结果与标准判定记录，按 PRD 中性值 50 分计入",
+            ),
+            self._stage_dimension(rows, stage_name),
+            self._defect_dimension(rows, category_name, material_name, party_name, defect_level),
+        )
+        score = round(sum(item.score * item.weight / 100 for item in dimensions), 2)
+        return DatabaseRiskScore(score=score, level=classify_score(score), dimensions=dimensions)
+
+    async def _load_rows(self) -> list[dict[str, Any]]:
+        start = self.as_of - timedelta(days=365 * 3)
+        statement = select(
+            AnomalyCase.date,
+            AnomalyCase.product_name,
+            AnomalyCase.product_category,
+            AnomalyCase.material_type,
+            AnomalyCase.supplier,
+            AnomalyCase.factory,
+            AnomalyCase.stage_new,
+            AnomalyCase.defect_level,
+        ).where(
+            AnomalyCase.date.is_not(None),
+            AnomalyCase.date >= start,
+            AnomalyCase.date <= self.as_of,
+        )
+        return [dict(row) for row in (await self.session.execute(statement)).mappings().all()]
+
+    def _supplier_dimension(self, rows: list[dict[str, Any]], party_name: str | None) -> DimensionScore:
+        weight = DEFAULT_WEIGHTS["供应商表现"]
+        if not party_name:
+            return DimensionScore(
+                name="供应商表现",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="未提供供应商/加工厂，按 PRD 缺失维度默认 50 分",
+            )
+        matched = [row for row in rows if party_matches(row, party_name)]
+        count = len(matched)
+        if count == 0:
+            base = 70.0
+        elif count <= 20:
+            base = 90.0
+        elif count <= 60:
+            base = 75.0
+        elif count <= 100:
+            base = 65.0
+        elif count <= 150:
+            base = 45.0
+        else:
+            base = 30.0
+        a_count = sum(1 for row in matched if normalize_defect(row.get("defect_level")) == "A")
+        base -= min(15, a_count * 5)
+        current_start = self.as_of - timedelta(days=183)
+        previous_start = current_start - timedelta(days=183)
+        current = sum(1 for row in matched if row.get("date") and row["date"] >= current_start)
+        previous = sum(
+            1
+            for row in matched
+            if row.get("date") and previous_start <= row["date"] < current_start
+        )
+        trend_adjustment = (
+            -10
+            if current > previous * 1.2 and current >= 3
+            else 5
+            if previous > current * 1.2
+            else 0
+        )
+        base = max(0.0, min(100.0, base + trend_adjustment))
+        alpha = min(1.5, 1.0 + count / 100 * 0.3)
+        score = min(100.0, round(base * alpha, 2))
+        return DimensionScore(
+            name="供应商表现",
+            score=score,
+            weight=weight,
+            status="database",
+            reason=f"近 3 年异常 {count} 条，基础分 {base:g}，α={alpha:.2f}",
+            sample_size=count,
+            raw_value=count,
+        )
+
+    @staticmethod
+    def _material_dimension(rows: list[dict[str, Any]], material: str | None) -> DimensionScore:
+        weight = DEFAULT_WEIGHTS["物料品类"]
+        if not material:
+            return DimensionScore(
+                name="物料品类",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="未提供包装材质，按 PRD 缺失维度默认 50 分",
+            )
+        counts: dict[str, int] = {}
+        for row in rows:
+            normalized = normalize_material(row.get("material_type"))
+            if normalized:
+                counts[normalized] = counts.get(normalized, 0) + 1
+        count = counts.get(material, 0)
+        maximum = max(counts.values(), default=0)
+        score = 100.0 if maximum == 0 else round(100 - count / maximum * 100, 2)
+        return DimensionScore(
+            name="物料品类",
+            score=score,
+            weight=weight,
+            status="database",
+            reason=f"物料「{material}」异常 {count} 条，相对最大物料异常 {maximum} 条归一化",
+            sample_size=count,
+            raw_value=count,
+        )
+
+    @staticmethod
+    def _category_dimension(rows: list[dict[str, Any]], category: str | None) -> DimensionScore:
+        weight = DEFAULT_WEIGHTS["产品品类"]
+        if not category:
+            return DimensionScore(
+                name="产品品类",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="无法从产品名称归类，按 PRD 缺失维度默认 50 分",
+            )
+        counts: dict[str, int] = {}
+        for row in rows:
+            normalized = normalize_category(row.get("product_category")) or normalize_category(row.get("product_name"))
+            if normalized:
+                counts[normalized] = counts.get(normalized, 0) + 1
+        count = counts.get(category, 0)
+        maximum = max(counts.values(), default=0)
+        score = 45.0 if count < 10 else round(100 - count / maximum * 100, 2) if maximum else 45.0
+        return DimensionScore(
+            name="产品品类",
+            score=score,
+            weight=weight,
+            status="database",
+            reason=(
+                f"品类「{category}」异常 {count} 条"
+                + (
+                    "，样本少于 10 条按 45 分处理"
+                    if count < 10
+                    else "，按品类最大异常数归一化"
+                )
+            ),
+            sample_size=count,
+            raw_value=count,
+        )
+
+    @staticmethod
+    def _stage_dimension(rows: list[dict[str, Any]], stage: str | None) -> DimensionScore:
+        weight = DEFAULT_WEIGHTS["节点阶段"]
+        if not stage:
+            return DimensionScore(
+                name="节点阶段",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="未提供产品阶段，按 PRD 缺失维度默认 50 分",
+            )
+        total = len(rows)
+        count = sum(1 for row in rows if normalize_stage(row.get("stage_new")) == stage)
+        share = count / total if total else 0
+        if share > 0.5:
+            score = 50.0
+        elif share >= 0.3:
+            score = 60.0
+        elif share >= 0.1:
+            score = 75.0
+        elif share >= 0.01:
+            score = 80.0
+        else:
+            score = 90.0
+        return DimensionScore(
+            name="节点阶段",
+            score=score,
+            weight=weight,
+            status="database",
+            reason=f"阶段「{stage}」异常占比 {share:.1%}，按阶段档位取分",
+            sample_size=count,
+            raw_value=share * 100,
+        )
+
+    @staticmethod
+    def _defect_dimension(
+        rows: list[dict[str, Any]],
+        category: str | None,
+        material: str | None,
+        party_name: str | None,
+        defect_level: str | None,
+    ) -> DimensionScore:
+        weight = DEFAULT_WEIGHTS["缺陷等级"]
+        if not category and not material and not party_name and not defect_level:
+            return DimensionScore(
+                name="缺陷等级",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="缺少品类、物料、供应商和缺陷等级组合，按 PRD 缺失维度默认 50 分",
+            )
+        matched = [
+            row
+            for row in rows
+            if (
+                not category
+                or normalize_category(row.get("product_category")) == category
+                or normalize_category(row.get("product_name")) == category
+            )
+            and (not material or normalize_material(row.get("material_type")) == material)
+            and (not party_name or party_matches(row, party_name))
+        ]
+        counts = {
+            level: sum(
+                1
+                for row in matched
+                if normalize_defect(row.get("defect_level")) == level
+            )
+            for level in ("A", "B", "C")
+        }
+        labeled = sum(counts.values())
+        if not labeled:
+            return DimensionScore(
+                name="缺陷等级",
+                score=50.0,
+                weight=weight,
+                status="default",
+                reason="当前组合没有可用的 A/B/C 缺陷记录，按 PRD 缺失维度默认 50 分",
+                sample_size=0,
+            )
+        penalty = counts["A"] * 40 + counts["B"] * 15 + counts["C"] * 5
+        score = max(0.0, 100.0 - penalty)
+        return DimensionScore(
+            name="缺陷等级",
+            score=score,
+            weight=weight,
+            status="database",
+            reason=f"组合缺陷记录 A={counts['A']}、B={counts['B']}、C={counts['C']}，按严重度扣分",
+            sample_size=labeled,
+            raw_value=penalty,
+        )
+
+
+def clean_value(value: object) -> str | None:
+    if value is None:
+        return None
+    cleaned = " ".join(str(value).strip().split())
+    return cleaned if cleaned and cleaned not in {"/", "-", "未填写", "未归类"} else None
+
+
+def normalize_defect(value: object) -> str | None:
+    cleaned = clean_value(value)
+    return cleaned.upper() if cleaned and cleaned.upper() in {"A", "B", "C"} else None
+
+
+def normalize_stage(value: object) -> str | None:
+    cleaned = clean_value(value)
+    if not cleaned:
+        return None
+    for stage in STAGES:
+        if stage in cleaned:
+            return stage
+    return None
+
+
+def normalize_material(value: object) -> str | None:
+    cleaned = clean_value(value)
+    if not cleaned:
+        return None
+    aliases = (
+        ("软管", "软管"),
+        ("泵头", "泵头"),
+        ("彩盒", "彩盒"),
+        ("标贴", "标贴"),
+        ("标签", "标贴"),
+        ("收缩膜", "收缩膜"),
+        ("瓶", "瓶子"),
+        ("PET", "瓶子"),
+        ("亚克力", "瓶子"),
+        ("盖", "盖子"),
+    )
+    for keyword, normalized in aliases:
+        if keyword.lower() in cleaned.lower():
+            return normalized
+    return cleaned
+
+
+def normalize_category(value: object) -> str | None:
+    cleaned = clean_value(value)
+    if not cleaned:
+        return None
+    aliases = (
+        (("洗发", "洗护", "洗发水", "洗发露", "洗发乳"), "洗发水/洗护类"),
+        (("沐浴", "沐浴露"), "沐浴露类"),
+        (("面霜", "乳液", "精华", "面膜", "护肤"), "面霜/护肤类"),
+        (("牙膏", "口腔"), "牙膏/口腔护理类"),
+        (("清洁剂", "清洁", "洗洁"), "清洁剂类"),
+        (("脱毛",), "脱毛类"),
+        (("彩妆", "口红", "粉底", "眼影"), "彩妆/唇部类"),
+        (("香氛", "香水", "止汗"), "香氛/止汗类"),
+    )
+    for keywords, normalized in aliases:
+        if any(keyword.lower() in cleaned.lower() for keyword in keywords):
+            return normalized
+    return None
+
+
+def party_matches(row: dict[str, Any], party_name: str) -> bool:
+    supplier = clean_value(row.get("supplier")) or ""
+    return party_name.lower() in supplier.lower()
+
+
+def classify_score(score: float) -> RiskLevel:
+    if score >= 80:
+        return RiskLevel.LOW
+    if score >= 60:
+        return RiskLevel.MODERATELY_LOW
+    if score >= 40:
+        return RiskLevel.MEDIUM
+    if score >= 20:
+        return RiskLevel.HIGH
+    return RiskLevel.VERY_HIGH

app/services/llm_client.py

@@ -1,17 +1,21 @@
-"""OpenAI-compatible LLM client for semantic extraction and streaming replies."""
+"""Provider-neutral LLM client with a configurable DeepSeek fallback."""
 
 import json
+import logging
 from collections.abc import AsyncGenerator
-from typing import Any
+from typing import Any, Literal
 
 import httpx
 
 from app.core.config import get_settings
 from app.core.exceptions import AppException
 
+logger = logging.getLogger("dieya_risk_prediction.llm_client")
+ProviderName = Literal["primary", "fallback"]
+
 
 class LLMClient:
-    """Small provider-neutral client using the OpenAI chat-completions protocol."""
+    """Call the primary provider and fail over to DeepSeek when necessary."""
 
     def __init__(self, transport: httpx.AsyncBaseTransport | None = None) -> None:
         settings = get_settings()
@@ -20,22 +24,56 @@ class LLMClient:
         self.model = settings.llm_model
         self.timeout = settings.llm_timeout_seconds
         self.temperature = settings.llm_temperature
+
+        self.fallback_api_key = (
+            settings.llm_fallback_api_key.get_secret_value() if settings.llm_fallback_api_key else ""
+        )
+        self.fallback_base_url = settings.llm_fallback_base_url.rstrip("/")
+        self.fallback_model = settings.llm_fallback_model
+        self.fallback_api_mode = settings.llm_fallback_api_mode
+        self.fallback_timeout = settings.llm_fallback_timeout_seconds
+        self.fallback_max_retries = settings.llm_fallback_max_retries
+        self.fallback_reasoning_effort = settings.llm_fallback_reasoning_effort
+        self.fallback_max_output_tokens = settings.llm_fallback_max_output_tokens
         self.transport = transport
 
     @property
     def configured(self) -> bool:
-        return bool(self.api_key)
+        return bool(self.api_key or self.fallback_api_key)
+
+    @property
+    def fallback_configured(self) -> bool:
+        return bool(self.fallback_api_key and self.fallback_base_url and self.fallback_model)
 
     @property
     def endpoint(self) -> str:
-        suffix = "/chat/completions"
+        return self._chat_endpoint(self.base_url)
-        return f"{self.base_url}{suffix}" if self.base_url.endswith("/v1") else f"{self.base_url}/v1{suffix}"
 
-    def _headers(self) -> dict[str, str]:
+    @staticmethod
-        if not self.configured:
+    def _chat_endpoint(base_url: str) -> str:
+        suffix = "/chat/completions"
+        return f"{base_url}{suffix}" if base_url.endswith("/v1") else f"{base_url}/v1{suffix}"
+
+    def _endpoint(self, provider: ProviderName) -> str:
+        if provider == "primary":
+            return self.endpoint
+        if self.fallback_api_mode == "responses":
+            return f"{self.fallback_base_url}/responses"
+        return self._chat_endpoint(self.fallback_base_url)
+
+    def _provider_config(self, provider: ProviderName) -> tuple[str, str, int]:
+        if provider == "primary":
+            return self.api_key, self.model, self.timeout
+        return self.fallback_api_key, self.fallback_model, self.fallback_timeout
+
+    def _headers(self, provider: ProviderName) -> dict[str, str]:
+        api_key, _, _ = self._provider_config(provider)
+        if not api_key:
+            if provider == "primary" and self.fallback_configured:
+                raise AppException("主 LLM 未配置 API Key，将尝试使用兜底模型", code=503, status_code=503)
             raise AppException("未配置 LLM_API_KEY，暂时无法使用 AI 对话", code=503, status_code=503)
         return {
-            "Authorization": f"Bearer {self.api_key}",
+            "Authorization": f"Bearer {api_key}",
             "Content-Type": "application/json",
             "Accept": "text/event-stream",
         }
@@ -45,14 +83,14 @@ class LLMClient:
         text: str,
         history: list[dict[str, str]] | None = None,
     ) -> dict[str, Any]:
-        """Extract structured product and intent fields before generating an answer."""
+        """Extract structured fields, retrying with DeepSeek when the primary fails."""
 
         messages = [
             {
                 "role": "system",
                 "content": (
                     "你是蝶雅产品风险预估系统的语义理解模块。"
-                    "只输出合法 JSON，不要输出 Markdown。抽取用户输入中的意图和风险评估参数。"
+                    "只输出合法 JSON，不要输出 Markdown。请输出 json object。抽取用户输入中的意图和风险评估参数。"
                     "intent 只能是 greet、general、assessment、supplier、material、cases、"
                     "standard、compliance、solution 之一。"
                     "字段必须包含 intent, risk_requested, category, specification, packaging, "
@@ -67,39 +105,136 @@ class LLMClient:
         ]
         messages.extend(history or [])
         messages.append({"role": "user", "content": text})
-        body = await self._request(messages, stream=False, json_mode=True)
+
         try:
-            content = body["choices"][0]["message"]["content"]
+            body = await self._request(messages, provider="primary", stream=False, json_mode=True)
-        except (KeyError, IndexError, TypeError) as exc:
+            content = self._response_text(body, provider="primary")
-            raise AppException("LLM 语义结果缺少 content", code=503, status_code=503) from exc
+            return self._parse_json(content)
-        if isinstance(content, list):
+        except Exception as primary_exc:  # noqa: BLE001 - fallback must cover provider and parse failures
-            content = "".join(str(part.get("text", "")) for part in content if isinstance(part, dict))
+            logger.warning(
-        return self._parse_json(str(content))
+                "LLM semantic request failed provider=primary model=%s error=%s",
+                self.model,
+                primary_exc.__class__.__name__,
+            )
+            if not self.fallback_configured:
+                raise
 
-    async def stream_reply(
+        try:
-        self,
+            body = await self._request(messages, provider="fallback", stream=False, json_mode=True)
-        messages: list[dict[str, str]],
+            content = self._response_text(body, provider="fallback")
-    ) -> AsyncGenerator[str, None]:
+            logger.info("LLM semantic request succeeded provider=fallback model=%s", self.fallback_model)
-        """Yield text deltas from an OpenAI-compatible streaming response."""
+            return self._parse_json(content)
-
+        except Exception as fallback_exc:  # noqa: BLE001 - normalize both provider failures
-        headers = self._headers()
+            logger.exception(
-        payload = {
+                "LLM semantic fallback failed provider=fallback model=%s",
-            "model": self.model,
+                self.fallback_model,
-            "messages": messages,
+            )
-            "temperature": self.temperature,
-            "stream": True,
-        }
-        timeout = httpx.Timeout(self.timeout, connect=min(15, self.timeout))
-        async with httpx.AsyncClient(timeout=timeout, transport=self.transport) as client:
-            async with client.stream("POST", self.endpoint, headers=headers, json=payload) as response:
-                if response.status_code >= 400:
-                    detail = await response.aread()
             raise AppException(
-                        f"LLM 请求失败（HTTP {response.status_code}）",
+                "主模型与兜底模型均调用失败，请检查网络、模型配置或额度",
                 code=503,
                 status_code=503,
-                        details=detail.decode("utf-8", errors="replace")[:500],
+            ) from fallback_exc
+
+    async def stream_reply(self, messages: list[dict[str, str]]) -> AsyncGenerator[str, None]:
+        """Yield reply deltas and fail over only before any primary text is emitted."""
+
+        emitted = False
+        try:
+            async for delta in self._stream_provider(messages, provider="primary"):
+                emitted = True
+                yield delta
+            return
+        except Exception as primary_exc:  # noqa: BLE001 - fallback is the public behavior
+            logger.warning(
+                "LLM stream failed provider=primary model=%s error=%s emitted=%s",
+                self.model,
+                primary_exc.__class__.__name__,
+                emitted,
             )
+            if emitted or not self.fallback_configured:
+                raise
+
+        logger.info("LLM stream switching provider=primary to provider=fallback model=%s", self.fallback_model)
+        async for delta in self._stream_provider(messages, provider="fallback"):
+            yield delta
+
+    async def generate_prevention_suggestions(self, context: dict[str, Any]) -> list[str]:
+        """Generate executable prevention actions from the current risk context."""
+
+        messages = [
+            {
+                "role": "system",
+                "content": (
+                    "你是蝶雅产品风险预估系统的预防方案生成模块。"
+                    "请根据提供的风险评估数据生成可执行的质量预防建议。"
+                    "只输出合法 json object，不要输出 Markdown，格式必须是 "
+                    '{"suggestions":["建议1","建议2"]}。'
+                    "输出 4 到 6 条建议，每条不超过 80 个汉字。"
+                    "建议必须优先针对分数较低的风险维度、历史异常频发项和命中的标准规则，"
+                    "明确检测或管控动作；可以包含责任角色、阶段和验收闭环。"
+                    "不得虚构数据库没有提供的案例、检测结果、法规条款或 RAG 检索结果。"
+                    "数据不足时明确写出需要补充的验证动作，不要把猜测写成事实。"
+                    "不要重复表达，不要输出标题、解释、风险评分或额外字段。"
+                ),
+            },
+            {
+                "role": "user",
+                "content": "以下是本次风险评估数据，请生成预防建议：\n" + json.dumps(
+                    context, ensure_ascii=False, separators=(",", ":")
+                ),
+            },
+        ]
+        try:
+            body = await self._request(messages, provider="primary", stream=False, json_mode=True)
+            raw = self._response_text(body, provider="primary")
+            return self._parse_suggestions(raw)
+        except Exception as primary_exc:  # noqa: BLE001 - prevention suggestions have the same failover policy
+            logger.warning(
+                "LLM prevention suggestion request failed provider=primary model=%s error=%s",
+                self.model,
+                primary_exc.__class__.__name__,
+            )
+            if not self.fallback_configured:
+                raise
+
+        try:
+            body = await self._request(messages, provider="fallback", stream=False, json_mode=True)
+            raw = self._response_text(body, provider="fallback")
+            suggestions = self._parse_suggestions(raw)
+            logger.info(
+                "LLM prevention suggestions generated provider=fallback model=%s count=%s",
+                self.fallback_model,
+                len(suggestions),
+            )
+            return suggestions
+        except Exception as fallback_exc:  # noqa: BLE001 - caller owns deterministic fallback behavior
+            logger.exception(
+                "LLM prevention suggestion fallback failed provider=fallback model=%s",
+                self.fallback_model,
+            )
+            raise AppException(
+                "主模型与兜底模型均无法生成预防建议",
+                code=503,
+                status_code=503,
+            ) from fallback_exc
+
+    async def _stream_provider(
+        self,
+        messages: list[dict[str, str]],
+        *,
+        provider: ProviderName,
+    ) -> AsyncGenerator[str, None]:
+        headers = self._headers(provider)
+        _, model, timeout_seconds = self._provider_config(provider)
+        is_responses = provider == "fallback" and self.fallback_api_mode == "responses"
+        payload = self._payload(messages, provider=provider, stream=True, json_mode=False)
+        timeout = httpx.Timeout(timeout_seconds, connect=min(15, timeout_seconds))
+        async with httpx.AsyncClient(timeout=timeout, transport=self.transport) as client:
+            try:
+                async with client.stream("POST", self._endpoint(provider), headers=headers, json=payload) as response:
+                    if response.status_code >= 400:
+                        detail = await response.aread()
+                        raise self._http_failure(response.status_code, detail)
                     async for line in response.aiter_lines():
                         if not line.startswith("data:"):
                             continue
@@ -110,44 +245,138 @@ class LLMClient:
                             chunk = json.loads(raw)
                         except json.JSONDecodeError:
                             continue
-                    choices = chunk.get("choices") or []
+                        delta = self._stream_delta(chunk, responses_api=is_responses)
-                    if not choices:
-                        continue
-                    delta = (choices[0].get("delta") or {}).get("content")
-                    if isinstance(delta, list):
-                        delta = "".join(
-                            str(part.get("text", "")) for part in delta if isinstance(part, dict)
-                        )
                         if delta:
-                        yield str(delta)
+                            yield delta
+            except AppException:
+                raise
+            except httpx.HTTPError as exc:
+                raise AppException(
+                    "LLM 流式连接失败",
+                    code=503,
+                    status_code=503,
+                ) from exc
+
+        logger.info("LLM stream completed provider=%s model=%s", provider, model)
 
     async def _request(
         self,
         messages: list[dict[str, str]],
         *,
+        provider: ProviderName,
         stream: bool,
         json_mode: bool = False,
     ) -> dict[str, Any]:
-        headers = self._headers()
+        headers = self._headers(provider)
+        payload = self._payload(messages, provider=provider, stream=stream, json_mode=json_mode)
+        _, model, timeout_seconds = self._provider_config(provider)
+        timeout = httpx.Timeout(timeout_seconds, connect=min(15, timeout_seconds))
+        logger.info(
+            "LLM request start provider=%s model=%s endpoint=%s stream=%s",
+            provider,
+            model,
+            self._endpoint(provider),
+            stream,
+        )
+        async with httpx.AsyncClient(timeout=timeout, transport=self.transport) as client:
+            try:
+                response = await client.post(self._endpoint(provider), headers=headers, json=payload)
+                if response.status_code >= 400:
+                    detail = await response.aread()
+                    raise self._http_failure(response.status_code, detail)
+            except AppException:
+                raise
+            except httpx.HTTPError as exc:
+                raise AppException("LLM 语义理解请求失败", code=503, status_code=503) from exc
+        try:
+            data = response.json()
+        except ValueError as exc:
+            raise AppException("LLM 返回格式无效", code=503, status_code=503) from exc
+        if not isinstance(data, dict):
+            raise AppException("LLM 返回格式无效", code=503, status_code=503)
+        logger.info("LLM request success provider=%s model=%s status=%s", provider, model, response.status_code)
+        return data
+
+    def _payload(
+        self,
+        messages: list[dict[str, str]],
+        *,
+        provider: ProviderName,
+        stream: bool,
+        json_mode: bool,
+    ) -> dict[str, Any]:
+        if provider == "fallback" and self.fallback_api_mode == "responses":
             payload: dict[str, Any] = {
-            "model": self.model,
+                "model": self.fallback_model,
+                "input": messages,
+                "stream": stream,
+                "max_output_tokens": self.fallback_max_output_tokens,
+            }
+            if self.fallback_reasoning_effort:
+                payload["reasoning_effort"] = self.fallback_reasoning_effort
+            return payload
+
+        _, model, _ = self._provider_config(provider)
+        payload = {
+            "model": model,
             "messages": messages,
             "temperature": self.temperature,
             "stream": stream,
         }
         if json_mode:
             payload["response_format"] = {"type": "json_object"}
-        timeout = httpx.Timeout(self.timeout, connect=min(15, self.timeout))
+        return payload
-        async with httpx.AsyncClient(timeout=timeout, transport=self.transport) as client:
+
+    @staticmethod
+    def _http_failure(status_code: int, detail: bytes) -> AppException:
+        return AppException(
+            f"LLM 请求失败（HTTP {status_code}）",
+            code=503,
+            status_code=503,
+            details=detail.decode("utf-8", errors="replace")[:500],
+        )
+
+    @staticmethod
+    def _response_text(body: dict[str, Any], *, provider: ProviderName) -> str:
+        if provider == "fallback":
+            output_text = body.get("output_text")
+            if isinstance(output_text, str) and output_text:
+                return output_text
+            parts: list[str] = []
+            for item in body.get("output") or []:
+                if not isinstance(item, dict):
+                    continue
+                content = item.get("content")
+                if isinstance(content, str):
+                    parts.append(content)
+                for part in content or []:
+                    if isinstance(part, dict) and part.get("type") in {"output_text", "text"}:
+                        parts.append(str(part.get("text") or ""))
+            if parts:
+                return "".join(parts)
+
         try:
-                response = await client.post(self.endpoint, headers=headers, json=payload)
+            content = body["choices"][0]["message"]["content"]
-                response.raise_for_status()
+        except (KeyError, IndexError, TypeError) as exc:
-            except httpx.HTTPError as exc:
+            raise AppException("LLM 语义结果缺少 content", code=503, status_code=503) from exc
-                raise AppException("LLM 语义理解请求失败", code=503, status_code=503) from exc
+        if isinstance(content, list):
-        try:
+            content = "".join(str(part.get("text", "")) for part in content if isinstance(part, dict))
-            return response.json()
+        return str(content)
-        except ValueError as exc:
+
-            raise AppException("LLM 返回格式无效", code=503, status_code=503) from exc
+    @staticmethod
+    def _stream_delta(chunk: dict[str, Any], *, responses_api: bool) -> str:
+        if responses_api:
+            event_type = chunk.get("type") or chunk.get("event")
+            if event_type in {"response.output_text.delta", "output_text.delta"}:
+                return str(chunk.get("delta") or "")
+            return ""
+        choices = chunk.get("choices") or []
+        if not choices:
+            return ""
+        delta = (choices[0].get("delta") or {}).get("content")
+        if isinstance(delta, list):
+            delta = "".join(str(part.get("text", "")) for part in delta if isinstance(part, dict))
+        return str(delta or "")
 
     @staticmethod
     def _parse_json(content: str) -> dict[str, Any]:
@@ -161,3 +390,28 @@ class LLMClient:
         if not isinstance(parsed, dict):
             raise AppException("LLM 语义结果格式无效", code=503, status_code=503)
         return parsed
+
+    @staticmethod
+    def _parse_suggestions(content: str) -> list[str]:
+        """Parse and normalize the model's suggestion-only JSON response."""
+
+        cleaned = content.strip()
+        if cleaned.startswith("```"):
+            cleaned = cleaned.removeprefix("```").removeprefix("json").removesuffix("```").strip()
+        try:
+            parsed = json.loads(cleaned)
+        except json.JSONDecodeError as exc:
+            raise AppException("LLM 预防建议不是合法 JSON", code=503, status_code=503) from exc
+        values = parsed.get("suggestions") if isinstance(parsed, dict) else parsed
+        if not isinstance(values, list):
+            raise AppException("LLM 预防建议格式无效", code=503, status_code=503)
+        suggestions: list[str] = []
+        for value in values:
+            if not isinstance(value, str):
+                continue
+            item = " ".join(value.strip().split())
+            if item and item not in suggestions:
+                suggestions.append(item[:100])
+        if not suggestions:
+            raise AppException("LLM 未返回有效预防建议", code=503, status_code=503)
+        return suggestions[:6]

app/services/risk_prediction.py

@@ -1,12 +1,17 @@
 """Risk prediction use-case orchestration."""
 
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.exceptions import AppException
 from app.domain.risk_engine import BaselineRiskEngine, IndicatorValue
 from app.schemas.risk_prediction import (
     ModelInfo,
     RiskAssessmentRequest,
     RiskAssessmentResult,
+    RiskDimensionScore,
     RiskFactorResult,
 )
+from app.services.database_risk_scoring import DatabaseRiskScoringService
 
 
 class RiskPredictionService:
@@ -14,10 +19,63 @@ class RiskPredictionService:
 
     @classmethod
     def model_info(cls) -> ModelInfo:
-        return ModelInfo(model_version=cls._engine.model_version)
+        return ModelInfo(model_version="database-seven-dimension-v1")
 
     @classmethod
-    def preview(cls, request: RiskAssessmentRequest) -> RiskAssessmentResult:
+    async def preview(
+        cls,
+        request: RiskAssessmentRequest,
+        session: AsyncSession | None = None,
+    ) -> RiskAssessmentResult:
+        if session and any(
+            value is not None
+            for value in (
+                request.product_category,
+                request.material_type,
+                request.supplier,
+                request.factory,
+                request.stage_new,
+                request.defect_level,
+            )
+        ):
+            database_score = await DatabaseRiskScoringService(session).score(
+                category=request.product_category,
+                material=request.material_type,
+                supplier=request.supplier,
+                factory=request.factory,
+                stage=request.stage_new,
+                defect_level=request.defect_level,
+            )
+            dimensions = [
+                RiskDimensionScore(
+                    dimension=item.name,
+                    score=item.score,
+                    weight=item.weight,
+                    status=item.status,
+                    reason=item.reason,
+                    sample_size=item.sample_size,
+                    raw_value=item.raw_value,
+                )
+                for item in database_score.dimensions
+            ]
+            return RiskAssessmentResult(
+                subject_id=request.subject_id,
+                score=database_score.score,
+                level=database_score.level,
+                model_version="database-seven-dimension-v1",
+                factors=[
+                    RiskFactorResult(
+                        name=item.name,
+                        value=round(100 - item.score, 2),
+                        weight=item.weight,
+                        contribution=round(item.score * item.weight / 100, 2),
+                    )
+                    for item in database_score.dimensions
+                ],
+                dimension_scores=dimensions,
+            )
+        if not request.indicators:
+            raise AppException("请填写产品评估参数", code=422, status_code=422)
         score = cls._engine.predict(
             [
                 IndicatorValue(name=item.name, value=item.value, weight=item.weight)
@@ -39,4 +97,3 @@ class RiskPredictionService:
                 for item in score.factors
             ],
         )
-

tests/test_conversations.py

@@ -1,5 +1,6 @@
 from datetime import UTC, datetime
 
+import pytest
 from app.models.entities import Conversation, ConversationMessage
 from app.services.conversations import ConversationService
 
@@ -119,3 +120,61 @@ def test_product_assessment_semantics_override_nlu_miss() -> None:
 
     assert normalized["intent"] == "assessment"
     assert normalized["risk_requested"] is True
+
+
+def test_short_known_supplier_follow_up_is_routed_to_supplier_card() -> None:
+    assert ConversationService._infer_supplier_follow_up(
+        "佳亿",
+        ["佳亿", "祥印", "瑞沣"],
+    ) == "佳亿"
+    assert ConversationService._infer_supplier_follow_up(
+        "这个供应商",
+        ["佳亿", "祥印"],
+    ) is None
+
+
+def test_unknown_standalone_name_is_treated_as_new_supplier() -> None:
+    assert ConversationService._infer_unknown_supplier_follow_up("德盛") == "德盛"
+    assert ConversationService._infer_unknown_supplier_follow_up("新供应商A") is None
+    assert ConversationService._infer_unknown_supplier_follow_up("测试供应商") is None
+    assert ConversationService._infer_unknown_supplier_follow_up("新供应商B") is None
+    assert ConversationService._infer_unknown_supplier_follow_up("洗发水") is None
+    assert ConversationService._infer_unknown_supplier_follow_up("帮我查德盛") is None
+    assert ConversationService._response_type("测试供应商", {"intent": "supplier"}) == "general"
+    assert ConversationService._response_type("新供应商A", {"intent": "supplier"}) == "general"
+
+
+@pytest.mark.asyncio
+async def test_any_product_without_history_uses_solution_card(monkeypatch: pytest.MonkeyPatch) -> None:
+    service = object.__new__(ConversationService)
+
+    async def no_history(_: str) -> int:
+        return 0
+
+    monkeypatch.setattr(service, "_count_category_anomalies", no_history)
+
+    response_type = ConversationService._response_type(
+        "评估防晒霜新品，50g，软管包装",
+        {"intent": "assessment", "risk_requested": True, "category": "防晒霜"},
+    )
+    result = await service._use_solution_card_without_history(
+        response_type,
+        "评估防晒霜新品，50g，软管包装",
+        {"intent": "assessment", "risk_requested": True, "category": "防晒霜"},
+    )
+
+    assert response_type == "risk_assessment"
+    assert result == "solution"
+
+
+@pytest.mark.asyncio
+async def test_unknown_product_without_category_uses_solution_card() -> None:
+    service = object.__new__(ConversationService)
+
+    result = await service._use_solution_card_without_history(
+        "risk_assessment",
+        "评估这个新产品的风险",
+        {"intent": "assessment", "risk_requested": True},
+    )
+
+    assert result == "solution"

tests/test_database_risk_scoring.py

@@ -0,0 +1,65 @@
+from datetime import date
+
+import pytest
+from app.services.database_risk_scoring import DatabaseRiskScoringService
+
+
+def _row(**values):
+    return {
+        "date": date(2026, 1, 10),
+        "product_name": "洗发水 500ml",
+        "product_category": None,
+        "material_type": "瓶子",
+        "supplier": "祥印",
+        "factory": "一一",
+        "stage_new": "第三阶段（产品开发）",
+        "defect_level": None,
+        **values,
+    }
+
+
+@pytest.mark.asyncio
+async def test_first_phase_score_uses_database_dimensions_and_prd_fallbacks() -> None:
+    service = DatabaseRiskScoringService(object(), as_of=date(2026, 9, 4))
+
+    async def load_rows():
+        return [
+            _row(),
+            _row(product_name="洗发水 250ml", defect_level="B"),
+        ]
+
+    service._load_rows = load_rows  # type: ignore[method-assign]
+    result = await service.score(
+        category="洗发水",
+        material="瓶子",
+        supplier="祥印",
+        stage="产品开发",
+    )
+
+    assert len(result.dimensions) == 7
+    assert sum(item.weight for item in result.dimensions) == 100
+    assert result.level.value == "medium"
+    assert result.dimensions[0].score == 95.57
+    assert result.dimensions[0].raw_value == 2
+    assert result.dimensions[1].score == 0
+    assert result.dimensions[2].score == 45
+    assert result.dimensions[3].status == "default"
+    assert result.dimensions[4].status == "default"
+    assert result.dimensions[5].score == 50
+    assert result.dimensions[6].score == 85
+    assert result.score == 52.39
+
+
+@pytest.mark.asyncio
+async def test_supplier_dimension_does_not_count_factory_records() -> None:
+    service = DatabaseRiskScoringService(object(), as_of=date(2026, 9, 4))
+
+    async def load_rows():
+        return [_row(supplier=None, factory="祥印")]
+
+    service._load_rows = load_rows  # type: ignore[method-assign]
+    result = await service.score(supplier="祥印")
+
+    supplier_dimension = result.dimensions[0]
+    assert supplier_dimension.score == 70
+    assert supplier_dimension.raw_value == 0

tests/test_llm_protocol.py

@@ -18,6 +18,9 @@ async def test_openai_compatible_semantic_and_stream_protocol(monkeypatch: pytes
         nonlocal calls
         calls += 1
         if calls == 1:
+            request_body = json.loads(request.content)
+            assert request_body["response_format"] == {"type": "json_object"}
+            assert "json" in request_body["messages"][0]["content"]
             return httpx.Response(
                 200,
                 json={
@@ -45,13 +48,86 @@ async def test_openai_compatible_semantic_and_stream_protocol(monkeypatch: pytes
 
     client = LLMClient(transport=httpx.MockTransport(handler))
     semantic = await client.extract_semantics("评估洗发水新品")
-    chunks = [
+    chunks = [chunk async for chunk in client.stream_reply([{"role": "user", "content": "继续分析"}])]
-        chunk
-        async for chunk in client.stream_reply([{"role": "user", "content": "继续分析"}])
-    ]
 
     assert semantic["intent"] == "assessment"
     assert semantic["confidence"] == 0.92
     assert "".join(chunks) == "第一段第二段"
     assert calls == 2
     get_settings.cache_clear()
+
+
+@pytest.mark.asyncio
+async def test_semantic_request_fails_over_to_deepseek_responses(monkeypatch: pytest.MonkeyPatch) -> None:
+    get_settings.cache_clear()
+    monkeypatch.setenv("RISK_LLM_API_KEY", "primary-key")
+    monkeypatch.setenv("RISK_LLM_BASE_URL", "https://primary.test/v1")
+    monkeypatch.setenv("RISK_LLM_MODEL", "primary-model")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_API_KEY", "fallback-key")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_BASE_URL", "https://fallback.test")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_MODEL", "deepseek-v4-pro")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_API_MODE", "responses")
+    calls: list[str] = []
+
+    async def handler(request: httpx.Request) -> httpx.Response:
+        calls.append(str(request.url))
+        if request.url.path.endswith("/chat/completions"):
+            return httpx.Response(503, json={"error": {"message": "primary unavailable"}})
+        body = json.loads(request.content)
+        assert body["model"] == "deepseek-v4-pro"
+        assert body["input"][-1]["content"] == "评估洗发水新品"
+        return httpx.Response(
+            200,
+            json={
+                "output_text": json.dumps(
+                    {"intent": "assessment", "risk_requested": True, "confidence": 0.8},
+                    ensure_ascii=False,
+                )
+            },
+        )
+
+    client = LLMClient(transport=httpx.MockTransport(handler))
+    result = await client.extract_semantics("评估洗发水新品")
+
+    assert result["intent"] == "assessment"
+    assert calls == [
+        "https://primary.test/v1/chat/completions",
+        "https://fallback.test/responses",
+    ]
+    get_settings.cache_clear()
+
+
+@pytest.mark.asyncio
+async def test_stream_request_fails_over_before_primary_emits_text(monkeypatch: pytest.MonkeyPatch) -> None:
+    get_settings.cache_clear()
+    monkeypatch.setenv("RISK_LLM_API_KEY", "primary-key")
+    monkeypatch.setenv("RISK_LLM_BASE_URL", "https://primary.test/v1")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_API_KEY", "fallback-key")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_BASE_URL", "https://fallback.test")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_MODEL", "deepseek-v4-pro")
+    monkeypatch.setenv("RISK_LLM_FALLBACK_API_MODE", "responses")
+    calls: list[str] = []
+
+    async def handler(request: httpx.Request) -> httpx.Response:
+        calls.append(str(request.url))
+        if request.url.path.endswith("/chat/completions"):
+            raise httpx.ConnectError("primary connection failed", request=request)
+        return httpx.Response(
+            200,
+            headers={"content-type": "text/event-stream"},
+            content=(
+                'data: {"type":"response.output_text.delta","delta":"兜底"}\n\n'
+                'data: {"type":"response.output_text.delta","delta":"回复"}\n\n'
+                "data: [DONE]\n\n"
+            ).encode(),
+        )
+
+    client = LLMClient(transport=httpx.MockTransport(handler))
+    chunks = [chunk async for chunk in client.stream_reply([{"role": "user", "content": "继续分析"}])]
+
+    assert "".join(chunks) == "兜底回复"
+    assert calls == [
+        "https://primary.test/v1/chat/completions",
+        "https://fallback.test/responses",
+    ]
+    get_settings.cache_clear()