Merge pull request 'fix：Jev 观测改为 evaluator，会话视图不再被 JSON 分类结果占位' (#14) from feat/0.3.0 into main

Reviewed-on: #14

app/clients/decision.py

@@ -63,8 +63,8 @@ class DecisionEngine:
         from app.observability.tracing import observe_decision, record_decision
 
         # 观测包在请求外面：成功才记结果，失败仍抛原来的 DecisionError
-        # generation 为 None 时 record_decision 什么都不做
-        with observe_decision(request) as generation:
+        # observation 为 None 时 record_decision 什么都不做
+        with observe_decision(request) as observation:
             try:
                 response = await client.system_one(
                     state=request.state,
@@ -93,7 +93,7 @@ class DecisionEngine:
             except TypeSafeError as error:
                 raise DecisionError("Jev SDK 调用失败", kind="sdk") from error
             parsed = self._parse_response(request, response)
-            record_decision(generation, parsed)
+            record_decision(observation, parsed)
             return parsed
 
     @staticmethod

app/observability/tracing.py

@@ -1,6 +1,6 @@
 """一次大屏图调用的 Langfuse 追踪
 
-不属于处理管道的任一步。把图调用收成可过滤的 trace，并把 Jev 记成当前节点下的 generation
+不属于处理管道的任一步。把图调用收成可过滤的 trace，并把 Jev 记成当前节点下的 evaluator
 每轮调用单独记录，首次请求与澄清恢复通过 session 关联，不修改业务状态
 """
 
@@ -100,7 +100,7 @@ def _screen_trace(payload, thread_id, session_id, user_id=None):
 
     # 明确回调所属项目，避免进程中存在多个客户端时跳过追踪
     handler = CallbackHandler(public_key=os.getenv("LANGFUSE_PUBLIC_KEY"))
-    # 在当前异步上下文执行回调，手动记录的 Jev generation 才会挂到所属节点。
+    # 在当前异步上下文执行回调，手动记录的 Jev 观测才会挂到所属节点。
     handler.run_inline = True
     resume = isinstance(payload, Command) and payload.resume is not None
     trace_input = (
@@ -162,9 +162,10 @@ def _run_output(state: dict[str, Any]) -> dict[str, Any]:
 
 @contextmanager
 def observe_decision(request: Any) -> Iterator[Any]:
-    """在当前节点下面记一次 Jev 调用，yield 的是 generation，未追踪时是 None
+    """在当前节点下面记一次 Jev 调用，yield 的是观测对象，未追踪时是 None
 
-    TypeSafe 不走 LangChain callback，所以要手记。类型必须是 generation，面板才能算模型和 token
+    TypeSafe 不走 LangChain callback，所以要手记。类型不能是 generation：会话视图只用 generation
+    拼对话，且只显示最浅的一条对话线，Jev 比 agent 的模型调用浅，用 generation 会把真正的对话挤掉
     只有图回调已经挂上当前 span 时才记。没有父 span 就跳过，避免 Jev 单独变成一条 trace
     这里不吞异常：失败时标 ERROR 后原样抛出，调用方的 DecisionError 类型不变
     """
@@ -177,7 +178,7 @@ def observe_decision(request: Any) -> Iterator[Any]:
     name = "classify-intent" if set(request.decisions) == {"intent"} else "run-decision"
     # 复用本应用已经选定的客户端
     with client.start_as_current_observation(
-        as_type="generation",
+        as_type="evaluator",
         name=name,
         model=model,
         input={
@@ -187,11 +188,11 @@ def observe_decision(request: Any) -> Iterator[Any]:
             },
             "state": request.state,
         },
-    ) as generation:
+    ) as observation:
         try:
-            yield generation
+            yield observation
         except Exception as error:
-            generation.update(
+            observation.update(
                 level="ERROR",
                 status_message=str(error),
                 metadata=_error_metadata(error),
@@ -199,29 +200,26 @@ def observe_decision(request: Any) -> Iterator[Any]:
             raise
 
 
-def record_decision(generation: Any, response: Any) -> None:
-    """把成功的 Jev 结果写回刚才那条 generation
+def record_decision(observation: Any, response: Any) -> None:
+    """把成功的 Jev 结果写回刚才那条观测
 
-    generation 为 None 表示这次没追踪，直接返回，不影响判断结果
+    observation 为 None 表示这次没追踪，直接返回，不影响判断结果
     只写选项、置信度和 token。不写原始 HTTP 响应
     """
-    if generation is None:
+    if observation is None:
         return
-    usage: dict[str, int] = {}
+    # evaluator 没有 model/usage 字段，模型和 token 只能放 metadata
+    metadata = {"model": response.model, "request_id": response.request_id or ""}
     if response.usage.input_tokens is not None:
-        usage["input_tokens"] = response.usage.input_tokens
+        metadata["input_tokens"] = str(response.usage.input_tokens)
     if response.usage.output_tokens is not None:
-        usage["output_tokens"] = response.usage.output_tokens
+        metadata["output_tokens"] = str(response.usage.output_tokens)
 
-    generation.update(
-        model=response.model,
+    observation.update(
         output={
             name: _result_summary(result) for name, result in response.results.items()
         },
-        usage_details=usage or None,
-        metadata={
-            "request_id": response.request_id or "",
-        },
+        metadata=metadata,
     )