fix(agent): enforce complete homework answers

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

@@ -122,6 +122,8 @@ class KnowledgeAgentService:
                     limit=cls._selection_limit(retrieval_question),
                     prefer_numbered_practice=practice_overview,
                 )
+                if practice_overview:
+                    selected.sort(key=lambda item: (item.knowledge.name, item.section.sort_order))
                 selected_contents = {
                     item.section.id: cls._protect_content(cls._read_complete_section(db, item.section))
                     for item in selected

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

@@ -116,7 +116,7 @@ class ModelClientService:
         return {
             "ok": True,
             "message": "Agent 调试完成",
-            "answer": answer[:4000],
+            "answer": answer,
             "prompt": rag_result.prompt,
             "modelName": model.model_name,
             "retrieveCount": len(rag_result.chunks),

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

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import re
 from dataclasses import dataclass
 
 from sqlalchemy import select
@@ -127,6 +128,8 @@ class PromptService:
         if not context:
             context = "本轮没有可靠的正式知识章节。对于一般常识可以谨慎回答；涉及课程、老师观点或公司业务时必须说明依据不足，不得编造。"
 
+        completeness_constraint = cls._list_completeness_constraint(question, chunks)
+
         messages: list[dict[str, str]] = [
             {
                 "role": "system",
@@ -163,13 +166,38 @@ class PromptService:
                 "不要求知识库中必须存在一个与‘某课程作业’完全同名的章节。"
                 "回答作业清单类问题时，应直接给出作业名称，并概括每项作业的用途或核心方向；"
                 "正文已经提供基本步骤时可以简要说明，不要把本轮已经取到的内容留到下一轮再问。"
-                "章节标题可以作为作业或练习名称，标题下的子章节是对应的正式说明。\n\n"
+                "章节标题可以作为作业或练习名称，标题下的子章节是对应的正式说明。"
+                f"{completeness_constraint}\n\n"
                 f"{context}"
             ),
         })
         messages.append({"role": "user", "content": question.strip()})
         return messages
 
+    @staticmethod
+    def _list_completeness_constraint(question: str, chunks: list[RetrievedChunk]) -> str:
+        is_practice_list = (
+            any(marker in question for marker in ("作业", "功课", "练习"))
+            and any(marker in question for marker in ("有哪些", "是什么", "包括什么", "都有什么", "列出", "汇总", "总结"))
+        )
+        numbered = [
+            chunk for chunk in chunks
+            if re.search(r"(?:练习|作业|功课)\s*[一二三四五六七八九十百\d]+", chunk.title)
+        ]
+        if not is_practice_list or not numbered:
+            return ""
+        titles = "\n".join(
+            f"{index}. {chunk.knowledge_name} / {chunk.title}"
+            for index, chunk in enumerate(numbered, start=1)
+        )
+        return (
+            "\n\n[作业清单完整性约束]\n"
+            f"本轮共召回 {len(numbered)} 个正式编号练习。回答必须覆盖下面全部 {len(numbered)} 项，"
+            "不得只挑选部分项目，不得合并成少数概念，不得新增清单外的作业。"
+            "应按知识库分组，逐项保留正式练习名称，并根据对应正文简要概括；即使回答较长也必须列完。\n"
+            f"必须覆盖的标题清单：\n{titles}"
+        )
+
     @staticmethod
     def render_messages(messages: list[dict[str, str]]) -> str:
         role_labels = {"system": "系统", "user": "用户", "assistant": "大本营答疑助手"}
@@ -180,8 +208,6 @@ class PromptService:
 
     @staticmethod
     def _clean_history_content(content: str) -> str:
-        import re
-
         return re.sub(r"<think[\s\S]*?</think\s*>", "", content, flags=re.IGNORECASE).strip()
 
     @classmethod

ai_knowledge_base_v2/apps/backend/tests/test_agent_runtime_config.py

@@ -10,7 +10,7 @@ from app.models.admin import Admin
 from app.models.ai_config import ModelConfig
 from app.schemas.admin import AgentRuntimeConfigSaveRequest
 from app.services.model_stream_service import _openai_stream_payload
-from app.services.model_service import _max_output_tokens
+from app.services.model_service import ModelClientService, _max_output_tokens
 from app.services.rag_service import RagResult
 
 
@@ -99,3 +99,18 @@ def test_saved_runtime_config_is_persisted_on_enabled_model():
         assert payload["top_p"] == 0.9
         assert payload["presence_penalty"] == 0.2
         assert payload["frequency_penalty"] == 0.4
+
+
+def test_agent_debug_does_not_truncate_long_answer(monkeypatch):
+    monkeypatch.setattr(
+        "app.services.model_service._call_configured_model",
+        lambda *_args, **_kwargs: "长" * 6000,
+    )
+
+    result = ModelClientService.debug_model(
+        _model(),
+        RagResult(question="测试", knowledge_scopes=[], chunks=[], prompt="测试", allow_general_knowledge=True),
+        {"max_token": 8192},
+    )
+
+    assert len(result["answer"]) == 6000

ai_knowledge_base_v2/apps/backend/tests/test_knowledge_agent.py

@@ -21,6 +21,7 @@ from app.models.knowledge import (
 )
 from app.models.chat import ChatMessage
 from app.services.knowledge_agent_service import Candidate, KnowledgeAgentService
+from app.services.rag_service import PromptService, RetrievedChunk
 
 
 def _database() -> Session:
@@ -227,6 +228,26 @@ def test_homework_overview_keeps_all_numbered_practices_before_selection():
     assert all(item.chunk.title.startswith("练习") for item in selected)
 
 
+def test_homework_overview_prompt_requires_every_recalled_title():
+    chunks = [
+        RetrievedChunk(
+            knowledge_id=1 if index <= 6 else 2,
+            knowledge_name="《原生·上》" if index <= 6 else "《原生·下》",
+            title=f"练习{index}：作业{index}",
+            content=f"作业{index}的正式说明",
+        )
+        for index in range(1, 15)
+    ]
+
+    constraint = PromptService._list_completeness_constraint("原生里的作业都有哪些", chunks)
+
+    assert "本轮共召回 14 个正式编号练习" in constraint
+    assert "必须覆盖下面全部 14 项" in constraint
+    assert "不得新增清单外的作业" in constraint
+    assert "练习1：作业1" in constraint
+    assert "练习14：作业14" in constraint
+
+
 def test_complete_section_includes_child_headings_but_stops_at_next_peer():
     with _database() as db:
         knowledge = _add_published_knowledge(db, knowledge_id=1, name="合一课程")