更新

configs/llm_profiles.json

@@ -0,0 +1,64 @@
+{
+  "profiles": [
+    {
+      "profile_id": "c8e185a64fa0",
+      "name": "glm-5",
+      "model": "glm-5",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:16:22.438297+00:00",
+      "updated_at": "2026-06-16T09:19:03.089865+00:00"
+    },
+    {
+      "profile_id": "54ddfe5aeb46",
+      "name": "deepseek-v4-pro",
+      "model": "deepseek-v4-pro",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:17:08.473904+00:00",
+      "updated_at": "2026-06-16T09:19:07.504082+00:00"
+    },
+    {
+      "profile_id": "25d035eef194",
+      "name": "qwen3.5-flash",
+      "model": "qwen3.5-flash",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:18:24.265619+00:00",
+      "updated_at": "2026-06-16T09:18:24.265619+00:00"
+    },
+    {
+      "profile_id": "ff1d0f417a5d",
+      "name": "deepseek-v4-flash",
+      "model": "deepseek-v4-flash",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:18:57.091549+00:00",
+      "updated_at": "2026-06-16T09:18:57.091549+00:00"
+    },
+    {
+      "profile_id": "5b04c49df9df",
+      "name": "text-embedding-v4",
+      "model": "text-embedding-v4",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:19:49.104004+00:00",
+      "updated_at": "2026-06-16T09:19:49.104004+00:00"
+    },
+    {
+      "profile_id": "b4f7c82859d5",
+      "name": "text-embedding-v3",
+      "model": "text-embedding-v3",
+      "base_url": "http://6.86.80.4:30080/v1",
+      "api_key": "sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8",
+      "timeout_seconds": 600,
+      "created_at": "2026-06-16T09:20:18.266540+00:00",
+      "updated_at": "2026-06-16T09:20:18.266540+00:00"
+    }
+  ]
+}

rag_eval/metrics/pipeline.py

@@ -94,6 +94,23 @@ class MetricPipeline:
                 reference=sample.ground_truth,
                 retrieved_contexts=sample.contexts,
             )
+        elif name == "noise_sensitivity":
+            coroutine = metric.ascore(
+                user_input=sample.question,
+                response=sample.answer,
+                reference=sample.ground_truth,
+                retrieved_contexts=sample.contexts,
+            )
+        elif name == "factual_correctness":
+            coroutine = metric.ascore(
+                response=sample.answer,
+                reference=sample.ground_truth,
+            )
+        elif name == "semantic_similarity":
+            coroutine = metric.ascore(
+                reference=sample.ground_truth,
+                response=sample.answer,
+            )
         else:
             raise ValueError(f"Unsupported metric: {name}")
 

rag_eval/metrics/registry.py

@@ -1,8 +1,13 @@
 """Supported metric names recognized by scenario validation and pipeline setup."""
 
 SUPPORTED_METRICS = {
+    # Core retrieval / generation metrics (always available).
     "faithfulness",
     "answer_relevancy",
     "context_recall",
     "context_precision",
+    # Robustness and end-to-end metrics (see 架构设计 §10.2).
+    "noise_sensitivity",      # 鲁棒性：对检索噪声的敏感度
+    "factual_correctness",    # 端到端：回答相对标准答案的事实正确性
+    "semantic_similarity",    # 端到端：回答与标准答案的语义相似度（embedding，无 LLM 调用）
 }

scenarios/offline/siemens-pdf-offline-smoke.yaml

@@ -9,6 +9,10 @@ metrics:
   - answer_relevancy
   - context_recall
   - context_precision
+  # 可选：鲁棒性 / 端到端指标（数据集已含 ground_truth，取消注释即可启用）
+  # - noise_sensitivity      # 鲁棒性：对检索噪声的敏感度
+  # - factual_correctness    # 端到端：事实正确性（相对标准答案）
+  # - semantic_similarity    # 端到端：语义相似度（embedding，无 LLM 调用）
 output_dir: ../../outputs/siemens-pdf-offline-smoke
 runtime:
   batch_size: 4

scenarios/online/sample-pdf-question-bank-online.yaml

@@ -1,7 +1,7 @@
 scenario_name: sample-pdf-question-bank-online
 mode: online
 dataset: ../../datasets/raw/generated/sample-pdf-question-bank.csv
-judge_model: deepseek-v4-pro
+judge_model: qwen3.5-flash
 embedding_model: text-embedding-v3
 metrics:
 - faithfulness
@@ -19,4 +19,4 @@ app_adapter:
   callable: apps.pdf_question_bank.adapter:run
   static_kwargs:
     source_chunks_path: ../../outputs/dataset-builds/sample-pdf-question-bank/latest/source_chunks.jsonl
-    model: deepseek-v4-flash
+    model: glm-5

scenarios/online/siemens-pdf-question-bank-online.yaml

@@ -1,28 +1,26 @@
 scenario_name: siemens-pdf-question-bank-online
 mode: online
 dataset: ../../datasets/raw/generated/siemens-pdf-question-bank.csv
-# judge_model: qwen3.5-flash
 judge_model: deepseek-v4-flash
 embedding_model: text-embedding-v3
-optimization_advisor: true      # 评测结束后自动生成优化建议报告
+optimization_advisor: true
 metrics:
 - faithfulness
 - answer_relevancy
 - context_recall
 - context_precision
-  # 已启用：鲁棒性 / 端到端指标（数据集已含 ground_truth）
+- noise_sensitivity
-  - noise_sensitivity        # 鲁棒性：对检索噪声的敏感度
+- factual_correctness
-  - factual_correctness      # 端到端：事实正确性（相对标准答案）
+- semantic_similarity
-  - semantic_similarity      # 端到端：语义相似度（embedding，无 LLM 调用）
 output_dir: ../../outputs/online/siemens-pdf-question-bank
 runtime:
-  batch_size: 4
+  batch_size: 3
-  app_concurrency: 4
+  app_concurrency: 3
-  metric_concurrency: 4
+  metric_concurrency: 3
-  max_samples: 50
+  max_samples: 10
 app_adapter:
   type: python
   callable: apps.siemens_pdf_qa.adapter:run
   static_kwargs:
     source_chunks_path: ../../outputs/dataset-builds/siemens-pdf-question-bank/latest/source_chunks.jsonl
-    model: deepseek-v4-flash
+    model: glm-5