Merge remote-tracking branch 'origin/qyqy_develop' into RM2_develop

tests/unit/api/test_customer_service_test_page.py

@@ -8,6 +8,6 @@ def test_customer_service_test_page_exposes_visitor_agent_flow() -> None:
         response = client.get("/customer-service-test/")
 
     assert response.status_code == 200
-    assert "奶龙基金智能助手" in response.text
+    assert "南方财富智能助手" in response.text
     assert "/api/v1/visitor-tokens" in response.text
     assert "/api/v1/agent-runs" in response.text

tests/unit/service/test_trade_service.py

@@ -25,7 +25,8 @@ from app.api.schemas.trading import (
     HoldingListResponse,
     PortfolioSummary,
 )
-from app.core.errors import FundQuoteUnavailableError
+from app.core.contracts import RequestContext
+from app.core.errors import FundQuoteUnavailableError, SuitabilityMismatchError
 from app.service.trade_service import TradeService, _FeeRule
 
 # ---------------------------------------------------------------------------
@@ -206,3 +207,28 @@ async def test_quote_freshness_only_blocks_trade_path() -> None:
     assert snapshot.price == Decimal("4.5000")
     with pytest.raises(FundQuoteUnavailableError, match="行情已过期"):
         await service._fetch_quote(product, enforce_freshness=True)  # type: ignore[arg-type]
+
+
+@pytest.mark.asyncio
+async def test_trade_suitability_uses_request_context_and_denies_mismatch() -> None:
+    evaluator = SimpleNamespace(
+        evaluate=AsyncMock(
+            return_value=SimpleNamespace(allowed=False, reason_code="RISK_LEVEL_MISMATCH")
+        )
+    )
+    service = TradeService(session=None, suitability_evaluator=evaluator)  # type: ignore[arg-type]
+    product = SimpleNamespace(
+        product_code="510500",
+        risk_level="R4",
+        risk_disclosure_required=1,
+        second_confirmation_required=0,
+    )
+    context = RequestContext(
+        user_id="9001", trace_id="trade-test", roles=("customer",), customer_ids=("9001",)
+    )
+
+    with pytest.raises(SuitabilityMismatchError, match="RISK_LEVEL_MISMATCH"):
+        await service._check_suitability(9001, product, context)  # type: ignore[arg-type]
+
+    evaluator.evaluate.assert_awaited_once()
+    assert evaluator.evaluate.await_args.kwargs["context"] is context

tests/unit/worker/test_knowledge_vector_worker.py

@@ -250,8 +250,8 @@ async def test_re_sync_after_content_update_carries_new_text() -> None:
         {"knowledge_id": "11"}
     )
 
-    assert writer.upserts[0]["fields"]["snippet"] != updated
-    assert writer.upserts[1]["fields"]["snippet"] == updated
+    assert writer.upserts[0]["fields"]["content"] != updated
+    assert writer.upserts[1]["fields"]["content"] == updated
 
 
 async def test_intent_keeps_contract_none_for_ordinary_knowledge() -> None:
@@ -320,11 +320,34 @@ async def test_dispatch_publishes_exactly_one_event_per_call() -> None:
     assert session.commits == 0
 
 
+#: 两套**都真实存在**的集合 schema（2026-09-13 实测 describe_collection）：
+#: 本机现库是主键 `doc_id` + 正文 `content`，另一套环境是 `knowledge_id` + `snippet`。
+#: 写侧靠 `describe_collection` 把**逻辑**字段名映射成这些**物理**名 ——
+#: 桩不提供它，探测就会失败、整条写入被拒（这正是修复前真机上的表现：
+#: 22 块知识全部 `RecoverableAgentError`，事件进死信，而检索侧看不出异常）。
+SCHEMA_DOC_ID = ("doc_id", "title", "content", "chapter", "section", "tags",
+                 "doc_no", "version", "visibility", "embedding")
+SCHEMA_KNOWLEDGE_ID = ("knowledge_id", "title", "snippet", "tags", "version", "embedding")
+
+
 class _FakeMilvusClient:
-    def __init__(self) -> None:
+    def __init__(self, fields: tuple[str, ...] = SCHEMA_DOC_ID) -> None:
         self.upserts: list[dict[str, Any]] = []
         self.deletes: list[dict[str, Any]] = []
         self.closed = False
+        self._fields = fields
+
+    async def describe_collection(self, *, collection_name: str) -> dict[str, Any]:
+        # 除向量字段外都带 `max_length`，与真实 VarChar 字段一致 —— 写侧靠它判断
+        # 「哪些是必填标量字段、这一行没给就要补空串」。
+        return {
+            "collection_name": collection_name,
+            "fields": [
+                {"name": name} if name == "embedding"
+                else {"name": name, "params": {"max_length": 1024}}
+                for name in self._fields
+            ],
+        }
 
     async def upsert(self, *, collection_name: str, data: list[dict[str, Any]]) -> None:
         self.upserts.append({"collection_name": collection_name, "data": data})
@@ -336,8 +359,22 @@ class _FakeMilvusClient:
         self.closed = True
 
 
-async def test_writer_upserts_one_row_keyed_by_knowledge_id() -> None:
-    client = _FakeMilvusClient()
+@pytest.mark.parametrize(
+    ("present_fields", "id_key", "text_key"),
+    [
+        (SCHEMA_DOC_ID, "doc_id", "content"),
+        (SCHEMA_KNOWLEDGE_ID, "knowledge_id", "snippet"),
+    ],
+)
+async def test_writer_maps_logical_fields_to_collection_schema(
+    present_fields: tuple[str, ...], id_key: str, text_key: str
+) -> None:
+    """逻辑字段名 → 该集合的物理名：**两套 schema 都必须能写**。
+
+    写侧此前硬编码 `knowledge_id`/`snippet`，遇到另一套 schema 时整条 upsert 直接失败；
+    而且这条路径只在真机上暴露（桩没有 describe_collection），所以长期没被测出来。
+    """
+    client = _FakeMilvusClient(present_fields)
     writer = MilvusKnowledgeWriter("http://milvus:19530")
     writer._client = client
 
@@ -345,20 +382,60 @@ async def test_writer_upserts_one_row_keyed_by_knowledge_id() -> None:
         collection="fin_faq_collection",
         knowledge_id="11",
         vector=[0.5] * 1024,
-        fields={"title": "标题", "snippet": "正文"},
+        fields={"title": "标题", "content": "正文"},
     )
 
-    assert client.upserts == [{
-        "collection_name": "fin_faq_collection",
-        "data": [{
-            "knowledge_id": "11",
-            # 向量字段名必须与集合 schema 一致（`embedding`，不是 `vector`）：
-            # 集合 `enable_dynamic_field=False`，写错键会让 upsert 直接失败。
-            "embedding": [0.5] * 1024,
-            "title": "标题",
-            "snippet": "正文",
-        }],
-    }]
+    assert client.upserts[0]["collection_name"] == "fin_faq_collection"
+    row = client.upserts[0]["data"][0]
+    assert row[id_key] == "11"
+    # 向量字段名必须与集合 schema 一致（`embedding`，不是 `vector`）：
+    # 集合 `enable_dynamic_field=False`，写错键会让 upsert 直接失败。
+    assert row["embedding"] == [0.5] * 1024
+    assert row["title"] == "标题"
+    assert row[text_key] == "正文"
+    # 集合有、这一行没给的 VARCHAR 字段补值：Milvus 对非 nullable 且无默认值的字段要求
+    # 必须提供，缺一个整条 upsert 就失败（实测缺 `chapter` 报 `Insert missed an field`）。
+    # 两套 schema 只有一套带这些字段，所以只断言"集合里有的那些"。
+    for optional_field in ("chapter", "section"):
+        if optional_field in present_fields:
+            assert row[optional_field] == ""
+    # `visibility` 例外：留空会被检索侧 `visibility == "public"` 的过滤把整行排除，
+    # 表现为"入库成功却一条都检索不到"（实测 22 块知识全部如此），所以按 FIELD_DEFAULTS 填。
+    if "visibility" in present_fields:
+        assert row["visibility"] == "public"
+
+
+async def test_writer_skips_fields_the_collection_does_not_have() -> None:
+    """集合没有的逻辑字段**跳过而不是报错**（另一套 schema 里没有 `intent` 字段）。"""
+    client = _FakeMilvusClient(SCHEMA_KNOWLEDGE_ID)
+    writer = MilvusKnowledgeWriter("http://milvus:19530")
+    writer._client = client
+
+    await writer.upsert(
+        collection="fin_faq_collection",
+        knowledge_id="11",
+        vector=[0.1] * 1024,
+        fields={"content": "正文", "intent": "chitchat"},
+    )
+
+    row = client.upserts[0]["data"][0]
+    assert "intent" not in row
+    assert row["snippet"] == "正文"
+
+
+async def test_writer_fails_closed_when_schema_is_unusable() -> None:
+    """探测不出必要字段时必须**报错**，不能静默写入检索不到的数据。"""
+    client = _FakeMilvusClient(("title", "embedding"))  # 既无主键、也无正文
+    writer = MilvusKnowledgeWriter("http://milvus:19530")
+    writer._client = client
+
+    with pytest.raises(RecoverableAgentError):
+        await writer.upsert(
+            collection="fin_faq_collection",
+            knowledge_id="11",
+            vector=[0.1] * 1024,
+            fields={"content": "正文"},
+        )
 
 
 async def test_writer_delete_targets_single_id() -> None:
@@ -387,6 +464,12 @@ async def test_writer_fails_closed_on_empty_input(kwargs: Mapping[str, Any]) ->
 
 async def test_writer_translates_backend_failure_to_recoverable() -> None:
     class Broken:
+        async def describe_collection(self, *, collection_name: str) -> dict[str, Any]:
+            return {
+                "collection_name": collection_name,
+                "fields": [{"name": name} for name in SCHEMA_DOC_ID],
+            }
+
         async def upsert(self, **_kwargs: Any) -> None:
             raise RuntimeError("milvus down")
 

tests/unit/worker/test_runtime_knowledge_wiring.py

@@ -62,6 +62,20 @@ class FakeMilvusClient:
         self.deletes: list[dict[str, Any]] = []
         self._log = log
 
+    async def describe_collection(self, *, collection_name: str) -> dict[str, Any]:
+        """写侧靠它把**逻辑**字段名映射成物理名；桩不提供会走"探测失败"分支。
+
+        这里给本机现库的 schema：主键 `doc_id`、正文 `content`
+        （另一套环境是 `knowledge_id`/`snippet`，两套都由 `knowledge_schema` 统一映射）。
+        """
+        return {
+            "collection_name": collection_name,
+            "fields": [{"name": name} for name in (
+                "doc_id", "title", "content", "chapter", "section",
+                "tags", "doc_no", "version", "visibility", "embedding",
+            )],
+        }
+
     async def upsert(self, *, collection_name: str, data: list[dict[str, Any]]) -> None:
         self._log.append("upsert")
         self.upserts.append({"collection_name": collection_name, "data": data})
@@ -180,11 +194,11 @@ async def test_dispatch_one_consumes_knowledge_sync_event(
     write = client.upserts[0]
     assert write["collection_name"] == "fin_faq_collection"
     row = write["data"][0]
-    assert row["knowledge_id"] == "11"
+    assert row["doc_id"] == "11"
     # 字段名必须是集合 schema 的 `embedding`：写成 `vector` 会让 upsert 失败、检索永远命中不到。
     assert VECTOR_FIELD == "embedding"
     assert len(row[VECTOR_FIELD]) == 1024
-    assert row["snippet"] == "交易日 15:00 前提交，T+1 确认份额。"
+    assert row["content"] == "交易日 15:00 前提交，T+1 确认份额。"
     # 嵌入的是知识行正文（不是事件 payload），端点走 customer_service/embedding。
     assert embedder.texts == ["交易日 15:00 前提交，T+1 确认份额。"]
     # 事件本身被标记为已投递，不再 pending。