Merge branch 'eval/elapsed-time-memoir-batch-chunk' into development

api/app/features/evaluation/judge_service.py

@@ -30,10 +30,7 @@ TJudgeOutput = TypeVar(
 _TURN_MAX = 768
 _CONV_JUDGE_JSON_MAX = 2048
 _CONV_HEADER = "【完整对话】（每轮以 `[Turn k]` 开头）\n\n"
-_MEMOIR_MAX = 12000
-_MEMOIR_JSON_MAX = 1536
 _COMPARE_STREAM_MAX = 6144
-_MEMOIR_EVIDENCE_MAX = 12000
 
 
 def _eval_judge_prompt_char_pool_for_context(context_window_tokens: int) -> int:
@@ -251,10 +248,12 @@ def _build_memoir_judge_prompt(
         "若存在 `lineage_tier=fallback` 或证据不足，须保守打分并写 `insufficient_evidence`。",
         "",
     ]
+    ev_cap = max(1, int(settings.eval_judge_memoir_evidence_max_chars))
+    body_cap = max(1, int(settings.eval_judge_memoir_body_max_chars))
     if notes:
         sections.extend(["【评审说明】", notes[:1200], ""])
     if source:
-        sections.extend(["【原始访谈/对话证据】", source[:_MEMOIR_EVIDENCE_MAX], ""])
+        sections.extend(["【原始访谈/对话证据】", source[:ev_cap], ""])
     else:
         sections.extend(
             [
@@ -274,8 +273,8 @@ def _build_memoir_judge_prompt(
             ]
         )
     if reference:
-        sections.extend(["【参考基线/导出成稿】", reference[:_MEMOIR_EVIDENCE_MAX], ""])
-    sections.extend(["【当前回忆录正文】", memoir_markdown[:_MEMOIR_MAX]])
+        sections.extend(["【参考基线/导出成稿】", reference[:ev_cap], ""])
+    sections.extend(["【当前回忆录正文】", memoir_markdown[:body_cap]])
     return "\n".join(sections)
 
 
@@ -493,11 +492,15 @@ class EvalJudgeService:
                 self._llm,
                 prompt,
                 MemoirJudgeOutput,
-                max_tokens=_MEMOIR_JSON_MAX,
+                max_tokens=max(512, int(settings.eval_judge_memoir_completion_max_tokens)),
                 agent="EvalJudgeService.judge_memoir",
             )
             return JudgeCallResult(output=out)
         except LLMCallError as e:
             error = _judge_error_message(e)
-            logger.warning("memoir judge failed: {}", error)
+            # 回忆录评审在 INFO 也要可见（eval-web 排障）；非异常路径、不刷堆栈
+            logger.info(
+                "event=eval_memoir_judge_llm_call_failed agent=EvalJudgeService.judge_memoir msg={}",
+                error,
+            )
             return JudgeCallResult(output=None, error=error)