fix: preserve agent debug conversation context

ai_knowledge_base_v2/apps/backend/app/schemas/admin.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from datetime import datetime
+from typing import Literal
 
 from pydantic import BaseModel, Field
 
@@ -77,12 +78,18 @@ class PromptSaveRequest(BaseModel):
     promptContent: str = Field(min_length=1)
 
 
+class AgentDebugHistoryMessage(BaseModel):
+    role: Literal["user", "assistant"]
+    content: str = Field(min_length=1, max_length=20000)
+
+
 class AgentDebugRequest(BaseModel):
     promptContent: str = Field(min_length=1)
     modelId: int = Field(gt=0)
     knowledgeIds: list[int] = Field(default_factory=list)
     knowledgeVersions: dict[int, int] = Field(default_factory=dict)
     question: str = Field(min_length=1, max_length=2000)
+    history: list[AgentDebugHistoryMessage] = Field(default_factory=list, max_length=100)
     temperature: float | None = Field(default=None, ge=0, le=2)
     topP: float | None = Field(default=None, ge=0, le=1)
     topK: int | None = Field(default=None, ge=1, le=1000)

ai_knowledge_base_v2/apps/backend/app/services/agent_debug_service.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import AsyncIterator
+from types import SimpleNamespace
 
 from sqlalchemy.orm import Session
 
@@ -12,31 +13,33 @@ from app.services.admin_service import OperationLogService
 from app.services.knowledge_agent_service import KnowledgeAgentService
 from app.services.model_stream_service import ModelStreamService
 from app.services.reasoning_policy_service import ReasoningPolicyService
-from app.services.rag_service import PromptService, RagResult
+from app.services.rag_service import RagResult
 
 
 class AgentDebugService:
     @staticmethod
     async def build_result(db: Session, payload: AgentDebugRequest) -> RagResult:
+        history = [
+            SimpleNamespace(id=index + 1, role=item.role, content=item.content)
+            for index, item in enumerate(payload.history)
+        ]
         rag_result = await KnowledgeAgentService.build_result(
             db,
             question=payload.question,
+            history=history,
             version_overrides=payload.knowledgeVersions,
             preview_knowledge_ids=payload.knowledgeIds,
+            prompt_override=payload.promptContent,
         )
-        debug_messages = [
-            {"role": "system", "content": payload.promptContent.strip()},
-            *(rag_result.messages or []),
-        ]
         return RagResult(
             question=rag_result.question,
             knowledge_scopes=rag_result.knowledge_scopes,
             chunks=rag_result.chunks,
-            prompt=PromptService.render_messages(debug_messages),
+            prompt=rag_result.prompt,
             allow_general_knowledge=rag_result.allow_general_knowledge,
             retrieval_log_id=rag_result.retrieval_log_id,
             tool_trace=rag_result.tool_trace,
-            messages=debug_messages,
+            messages=rag_result.messages,
         )
 
     @staticmethod

ai_knowledge_base_v2/apps/backend/app/services/knowledge_agent_service.py

@@ -83,6 +83,7 @@ class KnowledgeAgentService:
         version_overrides: dict[int, int] | None = None,
         preview_knowledge_ids: list[int] | None = None,
         context_trace: list[dict] | None = None,
+        prompt_override: str | None = None,
     ) -> RagResult:
         started = perf_counter()
         catalog = cls.get_knowledge_catalog(
@@ -205,6 +206,7 @@ class KnowledgeAgentService:
                 history,
                 session_summary,
                 summary_up_to_message_id,
+                prompt_override=prompt_override,
             )
             return RagResult(
                 question=question,
@@ -309,15 +311,23 @@ class KnowledgeAgentService:
         text = question.strip()
         if len(text) <= 16:
             return True
+        if re.search(r"(上面|前面|之前|刚才|此前|前一个|原来).{0,8}(问题|内容|回答|练习|流程)", text):
+            return True
+        if re.search(r"(再|重新|继续).{0,6}(回答|解释|说明|分析|看看)", text):
+            return True
         return bool(re.search(r"(这个|那个|上述|上面|前面|其中|第二|继续|然后呢|怎么办|怎么做|什么意思|为什么)$", text))
 
     @staticmethod
     def _fallback_rewrite(question: str, history) -> str:
-        previous_user = next(
-            (message.content.strip() for message in reversed(history) if message.role == "user" and message.content.strip()),
-            "",
-        )
-        return f"关于“{previous_user}”的追问：{question.strip()}" if previous_user else question.strip()
+        previous_users = [
+            message.content.strip()
+            for message in history
+            if message.role == "user" and message.content.strip()
+        ][-3:]
+        if not previous_users:
+            return question.strip()
+        context = "\n".join(f"历史问题{index + 1}：{content}" for index, content in enumerate(previous_users))
+        return f"结合以下历史问题回答当前追问：\n{context}\n当前追问：{question.strip()}"
 
     @staticmethod
     def get_knowledge_catalog(

ai_knowledge_base_v2/apps/backend/app/services/rag_service.py

@@ -118,8 +118,9 @@ class PromptService:
         history: list[ChatMessage] | None = None,
         session_summary: str | None = None,
         summary_up_to_message_id: int | None = None,
+        prompt_override: str | None = None,
     ) -> list[dict[str, str]]:
-        prompt = cls._load_active_prompt(db)
+        prompt = prompt_override.strip() if prompt_override and prompt_override.strip() else cls._load_active_prompt(db)
 
         # 知识库上下文
         context = "\n\n".join(cls._format_chunk(index, chunk) for index, chunk in enumerate(chunks, start=1))