fix: 修复会话上下文记忆链路

ai_knowledge_base_v2/apps/backend/tests/test_chat_context.py

@@ -0,0 +1,164 @@
+from __future__ import annotations
+
+from sqlalchemy import create_engine
+from sqlalchemy.orm import Session
+from sqlalchemy.pool import StaticPool
+
+from app.models import Base
+from app.models.ai_config import ModelConfig, SystemConfig
+from app.models.chat import ChatMessage, ChatSession
+from app.services.chat_context_service import ChatContextService
+from app.services.model_service import (
+    ModelClientService,
+    _fit_source_messages,
+    _max_output_tokens,
+    _openai_messages,
+    _rough_token_count,
+    _system_and_turn_messages,
+)
+from app.services.rag_service import PromptService, RagResult
+
+
+def _database(context_count: str = "10") -> Session:
+    engine = create_engine("sqlite:///:memory:", connect_args={"check_same_thread": False}, poolclass=StaticPool)
+    Base.metadata.create_all(engine)
+    db = Session(engine)
+    db.add(SystemConfig(id=1, config_key="chat_context_message_count", config_value=context_count))
+    db.commit()
+    return db
+
+
+def _messages(count: int) -> list[ChatMessage]:
+    result = []
+    for index in range(1, count + 1):
+        result.append(
+            ChatMessage(
+                id=index,
+                session_id=1,
+                user_id=1,
+                role="user" if index % 2 else "assistant",
+                content=f"消息{index}",
+            )
+        )
+    return result
+
+
+def test_context_message_count_is_read_at_runtime_and_zero_disables_all_memory():
+    with _database("2") as db:
+        history = _messages(4)
+        messages = PromptService.build_messages(db, "现在的问题", [], history, "旧摘要")
+        assert [item["content"] for item in messages if item["role"] in {"user", "assistant"}] == [
+            "消息3",
+            "消息4",
+            "现在的问题",
+        ]
+        assert any("旧摘要" in item["content"] for item in messages)
+
+        config = db.query(SystemConfig).filter_by(config_key="chat_context_message_count").one()
+        config.config_value = "0"
+        db.commit()
+        messages = PromptService.build_messages(db, "现在的问题", [], history, "旧摘要")
+        assert [item["content"] for item in messages if item["role"] in {"user", "assistant"}] == ["现在的问题"]
+        assert all("旧摘要" not in item["content"] for item in messages)
+
+
+def test_messages_leaving_window_are_summarized_immediately_and_summary_is_rewritten(monkeypatch):
+    with _database("4") as db:
+        session = ChatSession(id=1, user_id=1, title="测试", summary="已有摘要", message_count=6)
+        captured = {}
+
+        def summarize(_db, messages_text, previous_summary=None):
+            captured["messages"] = messages_text
+            captured["previous"] = previous_summary
+            return "覆盖后的完整摘要"
+
+        monkeypatch.setattr(ModelClientService, "summarize_or_raise", summarize)
+        result = ChatContextService.update_summary(db, session, _messages(6))
+
+        assert captured == {"messages": "用户：消息1\n大本营答疑助手：消息2", "previous": "已有摘要"}
+        assert session.summary == "覆盖后的完整摘要"
+        assert session.summary_up_to_message_id == 2
+        assert result.trace["status"] == "success"
+
+
+def test_messages_already_covered_by_summary_are_not_sent_twice_after_limit_increases():
+    with _database("4") as db:
+        messages = PromptService.build_messages(
+            db,
+            "现在的问题",
+            [],
+            _messages(4),
+            "已经覆盖消息1到消息3的摘要",
+            summary_up_to_message_id=3,
+        )
+        turns = [item["content"] for item in messages if item["role"] in {"user", "assistant"}]
+        assert turns == ["消息4", "现在的问题"]
+
+
+def test_summary_failure_is_visible_and_does_not_break_chat(monkeypatch, caplog):
+    with _database("2") as db:
+        session = ChatSession(id=9, user_id=1, title="测试", summary=None, message_count=4)
+
+        def fail(*_args, **_kwargs):
+            raise RuntimeError("摘要服务不可用")
+
+        monkeypatch.setattr(ModelClientService, "summarize_or_raise", fail)
+        result = ChatContextService.update_summary(db, session, _messages(4))
+
+        assert session.summary is None
+        assert result.trace["status"] == "fallback"
+        assert result.trace["error"] == "摘要服务不可用"
+        assert "Chat history summary failed" in caplog.text
+
+
+def test_provider_payload_uses_native_roles_and_normalizes_leading_assistant():
+    rag = RagResult(
+        question="当前问题",
+        knowledge_scopes=[],
+        chunks=[],
+        prompt="兼容文本",
+        messages=[
+            {"role": "system", "content": "系统规则"},
+            {"role": "assistant", "content": "窗口从助手消息开始"},
+            {"role": "user", "content": "历史问题"},
+            {"role": "assistant", "content": "历史回答"},
+            {"role": "system", "content": "知识上下文"},
+            {"role": "user", "content": "当前问题"},
+        ],
+    )
+
+    openai_messages = _openai_messages(rag)
+    system, turns = _system_and_turn_messages(rag)
+    assert [item["role"] for item in openai_messages] == ["system", "user", "assistant", "user"]
+    assert "窗口从助手消息开始" in system
+    assert "知识上下文" in system
+    assert turns[-1] == {"role": "user", "content": "当前问题"}
+
+
+def test_provider_messages_are_trimmed_to_configured_model_context_window():
+    rag = RagResult(
+        question="当前问题",
+        knowledge_scopes=[],
+        chunks=[],
+        prompt="兼容文本",
+        messages=[
+            {"role": "system", "content": "规则" * 500},
+            {"role": "user", "content": "很早的问题" * 200},
+            {"role": "assistant", "content": "很早的回答" * 200},
+            {"role": "system", "content": "知识" * 500},
+            {"role": "user", "content": "当前问题"},
+        ],
+    )
+    model = ModelConfig(
+        provider="test",
+        display_name="test",
+        api_type="openai_compatible",
+        model_name="test",
+        context_window=1000,
+        max_token=600,
+    )
+
+    fitted = _fit_source_messages(rag, model)
+    input_tokens = sum(_rough_token_count(item["content"]) for item in fitted)
+    assert input_tokens + _max_output_tokens(model) + 64 <= model.context_window
+    assert fitted[-1] == {"role": "user", "content": "当前问题"}

ai_knowledge_base_v2/apps/backend/tests/test_knowledge_agent.py

@@ -15,6 +15,7 @@ from app.models.knowledge import (
     KnowledgeSourceSnapshot,
     KnowledgeVersion,
 )
+from app.models.chat import ChatMessage
 from app.services.knowledge_agent_service import KnowledgeAgentService
 
 
@@ -118,7 +119,8 @@ def test_common_question_skips_knowledge_tools():
         result = asyncio.run(KnowledgeAgentService.build_result(db, question="今天北京天气怎么样？"))
         assert result.allow_general_knowledge is True
         assert result.chunks == []
-        assert result.tool_trace[1]["needKnowledge"] is False
+        decision = next(item for item in result.tool_trace if item["tool"] == "agent_decision")
+        assert decision["needKnowledge"] is False
 
 
 def test_course_question_without_catalog_does_not_fall_back_to_general_knowledge():
@@ -136,3 +138,19 @@ def test_course_question_searches_chunk_and_reads_parent_section():
         assert len(result.chunks) == 1
         assert "先稳定自己的焦虑" in result.chunks[0].content
         assert any(item["tool"] == "read_knowledge_sections" for item in result.tool_trace)
+
+
+def test_contextual_follow_up_is_rewritten_before_agent_decision():
+    with _database() as db:
+        history = [
+            ChatMessage(id=1, session_id=1, user_id=1, role="user", content="课程退款条件有哪些？"),
+            ChatMessage(id=2, session_id=1, user_id=1, role="assistant", content="第一种和第二种情况不同。"),
+        ]
+        result = asyncio.run(
+            KnowledgeAgentService.build_result(db, question="那第二种情况呢？", history=history)
+        )
+        rewrite = next(item for item in result.tool_trace if item["tool"] == "rewrite_contextual_question")
+        decision = next(item for item in result.tool_trace if item["tool"] == "agent_decision")
+
+        assert rewrite["rewrittenQuestion"] == "关于“课程退款条件有哪些？”的追问：那第二种情况呢？"
+        assert decision["request"]["question"] == rewrite["rewrittenQuestion"]