更新

- app.js: hash-based router (#runs / #new / #profiles / #report/{runId})
  - navigate() pushes history entries for back/forward support
  - _restoreSession() reads hash on load and popstate
  - sessionStorage fallback for same-tab refreshes
  - run-card highlights selected run (.run-card.selected)
- runner.js: use App.navigate() for report redirect; persist lastRunId to sessionStorage
- index.html: report nav button starts disabled (enabled on run select/restore)
- app.css: .run-card.selected with petrol border + ring

Co-Authored-By: Claude <noreply@anthropic.com>

.env.example

@@ -1,11 +1,22 @@
+# ===== LLM 连接配置（RAGAS 评测 + 生成） =====
+# 所有模型共用同一个 OpenAI 兼容 endpoint
+# 在 Web 控制台的「LLM 配置」页面可以保存多个命名配置，
+# 并在运行评估时按角色（Judge / Answer / Dataset）分别选择覆盖。
+
 OPENAI_API_KEY=your-api-key
 OPENAI_BASE_URL=http://6.86.80.4:30080/v1
+OPENAI_TIMEOUT_SECONDS=180
+
+# 默认评测模型（可在场景 YAML 或 Web 控制台 LLM 配置中覆盖）
 RAGAS_JUDGE_MODEL=deepseek-v4-flash
 RAGAS_EMBEDDING_MODEL=text-embedding-v3
+
+# 评估并发控制（启用 7 个指标时建议 RAGAS_METRIC_TIMEOUT_SECONDS=300）
 BATCH_SIZE=8
+RAGAS_METRIC_TIMEOUT_SECONDS=300
 
 
-# ===== 阿里云文档解析 =====
+# ===== 阿里云文档解析（dataset build 功能需要） =====
 ALIBABA_ACCESS_KEY_ID=
 ALIBABA_ACCESS_KEY_SECRET=
 ALIBABA_ENDPOINT=docmind-api.cn-hangzhou.aliyuncs.com
@@ -14,6 +25,8 @@ ALIYUN_PARSE_TIMEOUT_SECONDS=900
 ALIYUN_PARSE_LAYOUT_STEP_SIZE=50
 ALIYUN_LLM_ENHANCEMENT=true
 ALIYUN_ENHANCEMENT_MODE=VLM
-DOCUMENT_PARSE_ARTIFACT_PREFIX=artifacts
+DOCUMENT_PARSE_ARTIFACT_PREFIX=outputs/dataset-builds
 PARSER_FAILURE_MODE=fail
+
+# 生成题库时使用的模型（可在 Web 控制台 LLM 配置中按场景覆盖）
 DATASET_GENERATOR_MODEL=qwen3.6-plus

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"
+    }
+  ]
+}

docs/rag-eval-architecture.md

@@ -318,6 +318,10 @@ metrics:
   - answer_relevancy
   - context_recall
   - context_precision
+  # 可选：鲁棒性 / 端到端指标（需数据集含 ground_truth），完整列表见 §9.4
+  # - noise_sensitivity
+  # - factual_correctness
+  # - semantic_similarity
 output_dir: runs/legal-assistant-offline-baseline
 runtime:
   batch_size: 4
@@ -338,7 +342,7 @@ runtime:
 - `embedding_model`
   - 负责向量相关指标的模型
 - `metrics`
-  - 本次启用的指标列表
+  - 本次启用的指标列表（完整可选项与依赖见 §9.4）
 - `output_dir`
   - 本次运行结果输出目录
 - `runtime.batch_size`
@@ -399,6 +403,32 @@ app_adapter:
 - embedding model
 - 指标实例
 
+当前支持的指标（`rag_eval/metrics/registry.py` 中的 `SUPPORTED_METRICS`）：
+
+| 指标名 | 层面 | 依赖 |
+|---|---|---|
+| `faithfulness` | 生成 | judge model |
+| `answer_relevancy` | 生成 | judge model + embedding |
+| `context_recall` | 检索 | judge model + ground_truth |
+| `context_precision` | 检索 | judge model + ground_truth |
+| `noise_sensitivity` | 鲁棒性 | judge model + ground_truth |
+| `factual_correctness` | 端到端 | judge model + ground_truth |
+| `semantic_similarity` | 端到端 | embedding + ground_truth（无 LLM 调用） |
+
+后四项以 `ground_truth`（标准答案）为参照，数据集必须提供该字段。新增指标统一在 `registry.py` / `factory.py` / `pipeline.py` 三处对齐装配。
+
+**Optimization Advisor（§11 优化策略落地）：**
+
+评测结束后，若场景配置 `optimization_advisor: true`，则自动调用 `rag_eval/advisor/` 模块：
+- 规则引擎（`rules.py`）对 7 个指标各自设阈值，识别触发项并选取 top-3 低分样本
+- LLM 分析器（`llm_analyzer.py`）结合低分样本生成中文 Markdown 优化建议（复用 judge_model，失败自动降级为纯规则报告）
+- 写出层（`writer.py`）输出 `optimization_advice.md` 并打日志摘要
+
+```yaml
+# 场景配置示例
+optimization_advisor: true
+```
+
 ### 9.5 并发控制
 
 执行层负责并发上限，不把并发策略散落到各指标实现中。

docs/rag-eval-engine-flow.md

@@ -316,11 +316,21 @@ adapter 层的目标是：**把不同类型的目标应用，统一成同一套
 
 当前支持的指标包括：
 
+核心检索 / 生成指标（始终可用）：
+
 - `faithfulness`
 - `answer_relevancy`
 - `context_recall`
 - `context_precision`
 
+鲁棒性 / 端到端指标（架构设计 §10.2，需数据集含 `ground_truth`）：
+
+- `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（基于 embedding，无 LLM 调用）
+
+所有指标都通过同一套装配点接入：`registry.py`（校验白名单）、`factory.py`（实例化）、`pipeline.py`（`ascore` 入参分发），新增指标只需在这三处对齐即可。
+
 所以 metric pipeline 的职责可以总结为：
 
 **把标准样本转换成结构化评分结果。**
@@ -414,3 +424,39 @@ main.py
 - 可以把每次实验的资产稳定留住
 
 这也是它和一次性离线脚本的根本区别。
+
+---
+
+## 15. Optimization Advisor 链路
+
+相关代码：
+
+- `rag_eval/advisor/__init__.py` — 外部入口 `run_advisor()`
+- `rag_eval/advisor/rules.py` — 规则引擎（纯函数，无 LLM），7 条指标诊断规则
+- `rag_eval/advisor/llm_analyzer.py` — LLM 分析器（复用 judge_model llm 实例，失败自动降级）
+- `rag_eval/advisor/writer.py` — 写出 `optimization_advice.md` + 日志摘要
+
+Advisor 在 `write_run_artifacts()` 之后触发，仅当场景配置 `optimization_advisor: true` 时生效，默认关闭。
+
+执行链路：
+
+```text
+run_advisor(result, scenario, llm)
+  -> rules.diagnose(score_rows, metrics)     # 识别异常指标，选取 top-3 低分样本
+  -> llm_analyzer.analyze(diagnoses, llm)   # LLM 生成中文建议（失败自动降级为纯规则报告）
+  -> writer.write_advice(...)               # 写 optimization_advice.md + 日志摘要
+```
+
+输出产物追加在现有 run 目录：
+
+```text
+outputs/online/siemens-pdf-question-bank/<run_id>/
+  scenario.snapshot.yaml
+  scores.csv
+  invalid.csv
+  summary.md
+  metadata.json
+  optimization_advice.md    <- 新增（optimization_advisor: true 时生成）
+```
+
+规则引擎对 7 个指标各自设 warning / critical 双档阈值，`noise_sensitivity` 为"越低越好"（方向相反）。所有诊断均附带 top-3 低分样本，喂给 LLM 生成针对具体内容的中文建议。

docs/superpowers/specs/2026-06-16-optimization-advisor-design.md

@@ -0,0 +1,225 @@
+# 优化顾问模块设计 Spec
+
+- 日期：2026-06-16
+- 状态：已确认，进入实现。
+
+## 1. 目标
+
+在现有 RAG 评测流程结束后，新增一个**优化顾问模块**（Optimization Advisor），根据本次评测的多项指标分数与低分样本，自动诊断指标偏低的原因并给出针对性的优化建议，输出为中文 Markdown 报告 + 日志摘要。
+
+对应架构设计 §11（优化策略）：将"指标到动作的映射"（§11.2）从文档形式落地为代码自动执行。
+
+---
+
+## 2. 决策摘要
+
+| 决策点 | 选择 |
+|---|---|
+| 输出形式 | `optimization_advice.md`（文件）+ 控制台/日志摘要（双输出） |
+| 生成机制 | 规则引擎定位异常指标 → LLM 结合低分样本二次解读（两层） |
+| 触发方式 | YAML 场景文件显式声明 `optimization_advisor: true`，默认关闭 |
+| LLM 实例 | 复用 `build_models()` 已创建的 `llm` 实例，不重建 client |
+| 包位置 | `rag_eval/advisor/`（独立包，对外暴露 `run_advisor()` 单一入口） |
+
+---
+
+## 3. 架构
+
+### 3.1 执行链路
+
+```
+run_scenario()
+  → load_scenario()           # 读 YAML，解析 optimization_advisor 字段
+  → build_models()            # 已有：创建 llm, embeddings
+  → build_metric_pipeline()   # 已有
+  → Evaluator.evaluate()      # 已有：打分 → EvaluationResult
+  → write_run_artifacts()     # 已有：scores.csv / summary.md / ...
+  → run_advisor(              # 新增（3 行）
+        result, scenario, llm, artifact_paths
+    )
+      → rules.diagnose(score_rows)           # 规则引擎：返回 Diagnosis 列表
+      → llm_analyzer.analyze(diags, samples) # LLM：生成中文 Markdown 建议
+      → writer.write(advice, paths)          # 写文件 + 打日志
+```
+
+### 3.2 新增文件
+
+```
+rag_eval/advisor/
+  __init__.py          ← 暴露 run_advisor()，外部唯一入口
+  rules.py             ← 纯函数规则引擎，无 LLM，可单独单测
+  llm_analyzer.py      ← 接收 llm 实例 + 诊断结构 → 中文 Markdown
+  writer.py            ← 写 optimization_advice.md，打日志摘要
+```
+
+### 3.3 修改文件（最小改动）
+
+| 文件 | 改动 |
+|---|---|
+| `rag_eval/shared/models.py` | `Scenario` 加 `optimization_advisor: bool = False` 字段 |
+| `rag_eval/config/schema.py` | `ScenarioModel` 加同名字段 + 透传到 `Scenario` |
+| `rag_eval/config/loader.py` | 透传 `optimization_advisor` 到 `Scenario` 构造 |
+| `rag_eval/reporting/artifacts.py` | `RunArtifactPaths` 加 `advice_md: Path` 字段 + `build_artifact_paths()` 加赋值 |
+| `rag_eval/execution/runner.py` | `run_scenario()` 末尾：`build_models` 返回 llm 传入，条件调用 `run_advisor()` |
+
+### 3.4 输出产物
+
+```
+outputs/online/siemens-pdf-question-bank/<run_id>/
+  scenario.snapshot.yaml
+  scores.csv
+  invalid.csv
+  summary.md
+  metadata.json
+  optimization_advice.md    ← 新增（optimization_advisor: true 时生成）
+```
+
+---
+
+## 4. 规则引擎（rules.py）
+
+### 4.1 数据结构
+
+```python
+@dataclass
+class Diagnosis:
+    metric: str           # 指标名
+    mean_score: float     # 本次均值
+    threshold: float      # 警戒阈值
+    severity: str         # "warning" | "critical"
+    root_causes: list[str]  # 可能原因（来自架构设计 §11.2）
+    suggested_actions: list[str]  # 对应可调阶段
+    low_samples: list[dict]  # 分数最低的 N 条样本（含 question/answer/ground_truth）
+```
+
+### 4.2 七条指标诊断规则
+
+阈值参考 RAG 评测最佳实践，分 warning / critical 两档：
+
+| 指标 | warning | critical | 根因方向 | 对应优化阶段（§11.2） |
+|---|---|---|---|---|
+| `faithfulness` | < 0.7 | < 0.5 | 生成未严格基于检索片段 / 幻觉 | 生成 prompt grounding、开启校验 |
+| `answer_relevancy` | < 0.7 | < 0.5 | 回答偏离问题 / 格式冗余 | 查询改写、生成 prompt 格式 |
+| `context_recall` | < 0.7 | < 0.5 | 检索遗漏关键信息 | 多查询、问题分解、Step-back、加大过召回 |
+| `context_precision` | < 0.6 | < 0.4 | 检索引入过多噪声 / 排序差 | 后检索重排、压缩、相关性过滤 |
+| `noise_sensitivity` | > 0.3 | > 0.5 | 回答被噪声片段干扰（越低越好） | 后检索相关性过滤、重排 |
+| `factual_correctness` | < 0.6 | < 0.4 | 回答事实与标准答案偏差大 | 检索与生成综合优化 |
+| `semantic_similarity` | < 0.7 | < 0.5 | 回答语义与标准答案差距大 | 生成 prompt、检索质量 |
+
+> 注：`noise_sensitivity` 越低越好（0=完全不受噪声影响），其阈值方向与其余相反。
+
+### 4.3 低分样本选取
+
+每个触发诊断的指标，取该指标分数最低的 **top-3** 样本（排除 NaN）附入 `Diagnosis.low_samples`，字段包含 `sample_id / question / answer / ground_truth / <metric_score>`。
+
+---
+
+## 5. LLM 分析器（llm_analyzer.py）
+
+### 5.1 输入
+
+- `diagnoses: list[Diagnosis]` — 规则引擎输出（仅触发阈值的指标）
+- `llm` — 已有 RAGAS LLM 实例（scenario 的 judge_model）
+- `scenario_name: str` — 用于报告标题
+
+### 5.2 Prompt 设计
+
+使用**一次 LLM 调用**，把所有触发诊断的指标和低分样本一起发送：
+
+```
+你是一个 RAG 系统优化专家，正在分析西门子医疗 CT 文档问答系统的评测结果。
+请用中文撰写一份优化建议报告，格式为 Markdown。
+
+## 评测诊断摘要
+{for each diagnosis: 指标名、均值、阈值、可能原因、建议动作}
+
+## 低分样本示例
+{for each diagnosis: top-3 低分样本的 question / answer / ground_truth}
+
+## 要求
+1. 按指标分节（## 指标名），先解释"为什么低"，再给出"具体怎么改"
+2. "具体怎么改"要结合低分样本的具体内容，而不只是泛泛建议
+3. 最后写一节 ## 优先优化次序，按性价比排序（参考：不增加调用次数的优先）
+4. 语言简洁，面向工程师，不要废话
+```
+
+### 5.3 输出
+
+LLM 返回的 Markdown 字符串，直接写入 `optimization_advice.md`（在报告头部追加运行元信息）。
+
+### 5.4 失败降级
+
+LLM 调用失败（超时/异常）时：降级为**纯规则报告**（只输出规则引擎的诊断结构，不含 LLM 解读），文件照常写出，错误信息写入报告末尾，不阻断整个评测流程。
+
+---
+
+## 6. 写出层（writer.py）
+
+### 6.1 文件写出
+
+`optimization_advice.md` 结构：
+
+```markdown
+# 优化建议报告 — <scenario_name>
+
+- run_id: `<run_id>`
+- 生成时间: `<timestamp>`
+- judge_model: `<model>`
+
+---
+
+<LLM 生成的 Markdown 正文>
+```
+
+### 6.2 日志摘要
+
+`run_advisor()` 完成后向 `logger.info` 打印一条精简摘要（单行，适合 `run_eval.bat` 结束后一眼扫到）：
+
+```
+[advisor] 触发诊断 3 项: faithfulness(0.42, critical) context_recall(0.58, warning) noise_sensitivity(0.41, critical)
+[advisor] 优化建议已写出: outputs/online/.../optimization_advice.md
+```
+
+---
+
+## 7. YAML 配置
+
+场景文件新增一个顶层字段：
+
+```yaml
+optimization_advisor: true   # 默认 false；true 时评测结束后自动生成优化建议
+```
+
+后续若需精细配置（阈值覆盖、top-N 低分样本数），可扩展为：
+
+```yaml
+optimization_advisor:
+  enabled: true
+  top_low_samples: 3          # 每个指标取几条低分样本（默认 3）
+  # thresholds:               # 可选：覆盖默认阈值
+  #   faithfulness: 0.65
+```
+
+本轮实现仅支持 `optimization_advisor: true/false`，扩展接口预留但不实现。
+
+---
+
+## 8. 测试策略
+
+| 测试 | 文件 | 说明 |
+|---|---|---|
+| 规则引擎单测 | `tests/test_advisor_rules.py` | 纯函数，无 LLM，覆盖每条规则的 warning/critical 触发、NaN 跳过、low_samples 选取 |
+| writer 单测 | `tests/test_advisor_writer.py` | mock Diagnosis 列表，验证 md 文件写出格式和日志输出 |
+| 集成（可选） | 现有 `tests/test_online_eval.py` | 验证 `optimization_advisor: true` 场景下 advice_md 存在 |
+
+LLM 分析器不写单测（依赖网络），由集成场景覆盖。
+
+---
+
+## 9. 不覆盖（本轮边界）
+
+- 不支持跨版本对比分析（只分析本次 run）
+- 不支持批量场景聚合建议
+- 不建设 Web UI 展示
+- LLM 分析器 prompt 本轮不做多语言适配（直接中文）
+- advisor 阈值本轮硬编码在 `rules.py`，不从 YAML 读取

main.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 import argparse
+import logging
+from pathlib import Path
 
 from rag_eval.dataset_builder.runner import run_dataset_build
 from rag_eval.execution.runner import run_scenario
@@ -18,18 +20,33 @@ def parse_args() -> argparse.Namespace:
         "--dataset-build-config",
         help="Path to a YAML dataset build config file.",
     )
+    parser.add_argument(
+        "--log-file",
+        default=None,
+        help="Write evaluation logs to this file (in addition to stderr). "
+             "Example: logs/eval.log",
+    )
+    parser.add_argument(
+        "--log-level",
+        default="INFO",
+        choices=["DEBUG", "INFO", "WARNING", "ERROR"],
+        help="Logging verbosity level (default: INFO). Use DEBUG for per-metric detail.",
+    )
     return parser.parse_args()
 
 
 def main() -> None:
     """Dispatch the CLI call to the requested workflow."""
     args = parse_args()
+    log_level = getattr(logging, args.log_level.upper(), logging.INFO)
+    log_file = Path(args.log_file) if args.log_file else None
+
     if args.dataset_build_config:
         result = run_dataset_build(args.dataset_build_config)
         print(f"Completed dataset build: {result.artifact_paths.root_dir}")
         return
 
-    result = run_scenario(args.scenario)
+    result = run_scenario(args.scenario, log_file=log_file, log_level=log_level)
     print(f"Completed run: {result.scenario.output_dir}")
 
 

rag_eval/advisor/__init__.py

@@ -0,0 +1,67 @@
+"""Optimization advisor: rule-based diagnosis + LLM-powered recommendations."""
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Any
+
+from rag_eval.reporting.artifacts import build_artifact_paths
+from rag_eval.shared.models import EvaluationResult, Scenario
+
+from .llm_analyzer import analyze
+from .rules import Diagnosis, diagnose
+from .writer import write_advice
+
+logger = logging.getLogger("rag_eval.advisor")
+
+__all__ = ["run_advisor", "Diagnosis", "diagnose"]
+
+
+def run_advisor(
+    result: EvaluationResult,
+    scenario: Scenario,
+    llm: Any,
+) -> None:
+    """Run the full optimization advisor pipeline after an evaluation completes.
+
+    Skips silently if scenario.optimization_advisor is False.
+    Never raises — failures are logged as warnings, not exceptions.
+
+    Args:
+        result: Completed EvaluationResult from Evaluator.evaluate().
+        scenario: The resolved Scenario (provides metrics, judge_model, output_dir).
+        llm: Pre-built RAGAS LLM instance (from build_models()) for LLM analysis.
+    """
+    if not scenario.optimization_advisor:
+        return
+
+    logger.info("[advisor] starting optimization analysis  scenario=%s", scenario.scenario_name)
+
+    try:
+        artifact_paths = build_artifact_paths(scenario.output_dir, result.run_id)
+        if artifact_paths.advice_md is None:
+            logger.warning("[advisor] advice_md path not set in RunArtifactPaths — skipping")
+            return
+
+        diagnoses = diagnose(result.score_rows, scenario.metrics)
+        logger.info("[advisor] rule diagnosis complete: %d metric(s) triggered", len(diagnoses))
+
+        if diagnoses:
+            llm_markdown = asyncio.run(analyze(diagnoses, llm, scenario.scenario_name))
+        else:
+            llm_markdown = ""
+
+        write_advice(
+            diagnoses=diagnoses,
+            llm_markdown=llm_markdown,
+            advice_path=artifact_paths.advice_md,
+            scenario_name=scenario.scenario_name,
+            run_id=result.run_id,
+            judge_model=scenario.judge_model,
+        )
+
+    except Exception as exc:
+        logger.warning(
+            "[advisor] advisor failed (%s: %s) — evaluation result is unaffected",
+            type(exc).__name__, exc,
+        )

rag_eval/advisor/llm_analyzer.py

@@ -0,0 +1,100 @@
+"""LLM-powered analysis of rule diagnostics and low-score samples."""
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+from .rules import Diagnosis
+
+logger = logging.getLogger("rag_eval.advisor")
+
+_PROMPT_TEMPLATE = """\
+你是一个 RAG 系统优化专家，正在分析西门子医疗 CT 文档问答系统的评测结果。
+请用中文撰写一份优化建议报告，格式为 Markdown。
+
+## 评测诊断摘要
+
+{diagnosis_summary}
+
+## 低分样本示例
+
+{low_sample_text}
+
+## 报告要求
+
+1. 按指标分节（## 指标名  [severity]），先解释"为什么低"（结合低分样本具体分析），再给出"具体怎么改"
+2. "具体怎么改"要结合低分样本的实际内容，而不只是泛泛建议
+3. 最后写一节 **## 优先优化次序**，按性价比排序（不增加 LLM 调用次数的优化优先）
+4. 语言简洁，面向工程师，不要废话，不要重复列表内容
+
+只输出 Markdown 报告正文，不要任何前置说明。
+"""
+
+
+def _build_diagnosis_summary(diagnoses: list[Diagnosis]) -> str:
+    lines = []
+    for d in diagnoses:
+        direction = "（越低越好）" if d.metric == "noise_sensitivity" else ""
+        lines.append(
+            f"- **{d.metric}** {direction} 均值={d.mean_score:.4f}，"
+            f"阈值={d.threshold}，严重程度={d.severity}"
+        )
+        lines.append(f"  - 可能原因：{'; '.join(d.root_causes)}")
+        lines.append(f"  - 建议动作：{'; '.join(d.suggested_actions)}")
+    return "\n".join(lines)
+
+
+def _build_low_sample_text(diagnoses: list[Diagnosis]) -> str:
+    lines = []
+    for d in diagnoses:
+        if not d.low_samples:
+            continue
+        lines.append(f"### {d.metric} 低分样本（最多 3 条）")
+        for i, s in enumerate(d.low_samples, 1):
+            score = s.get(d.metric, "N/A")
+            lines.append(f"\n**样本 {i}**（分数={score}）")
+            lines.append(f"- 问题：{s.get('question', '')}")
+            lines.append(f"- 回答：{s.get('answer', '')[:300]}")
+            lines.append(f"- 标准答案：{s.get('ground_truth', '')[:200]}")
+    return "\n".join(lines)
+
+
+async def analyze(
+    diagnoses: list[Diagnosis],
+    llm: Any,
+    scenario_name: str,
+) -> str:
+    """Call the judge LLM to generate a Chinese optimization report.
+
+    Args:
+        diagnoses: Non-empty list of Diagnosis from rules.diagnose().
+        llm: RAGAS LLM wrapper (has .agenerate() method).
+        scenario_name: Used only for logging.
+
+    Returns:
+        LLM-generated Markdown string, or "" on failure (triggers writer fallback).
+    """
+    if not diagnoses:
+        return ""
+
+    diagnosis_summary = _build_diagnosis_summary(diagnoses)
+    low_sample_text = _build_low_sample_text(diagnoses)
+    prompt = _PROMPT_TEMPLATE.format(
+        diagnosis_summary=diagnosis_summary,
+        low_sample_text=low_sample_text,
+    )
+
+    try:
+        logger.info("[advisor] calling LLM for optimization analysis  scenario=%s", scenario_name)
+        from langchain_core.messages import HumanMessage
+        # Use the underlying langchain chat model directly (RAGAS LangchainLLMWrapper wraps BaseChatModel)
+        response = await llm.langchain_llm.ainvoke([HumanMessage(content=prompt)])
+        text = response.content.strip()
+        logger.info("[advisor] LLM analysis complete  chars=%d", len(text))
+        return text
+    except Exception as exc:
+        logger.warning(
+            "[advisor] LLM analysis failed (%s: %s) — falling back to rule report",
+            type(exc).__name__, exc,
+        )
+        return ""

rag_eval/advisor/rules.py

@@ -0,0 +1,236 @@
+"""Rule-based diagnostic engine for RAG evaluation metric scores."""
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class MetricRule:
+    """Threshold configuration and diagnostic text for one metric."""
+    warning_threshold: float
+    critical_threshold: float
+    higher_is_better: bool  # False for noise_sensitivity
+    root_causes: list[str]
+    suggested_actions: list[str]
+
+
+METRIC_RULES: dict[str, MetricRule] = {
+    "faithfulness": MetricRule(
+        warning_threshold=0.7,
+        critical_threshold=0.5,
+        higher_is_better=True,
+        root_causes=[
+            "生成回答包含检索片段中不支持的陈述（幻觉）",
+            "生成阶段未严格遵循 grounding 约束",
+            "校验阶段未开启或未生效",
+        ],
+        suggested_actions=[
+            "强化生成 prompt 的 grounding 约束（'只依据参考资料作答'）",
+            "开启校验阶段（validation: by_scenario）",
+            "检查低分样本中模型是否引用了片段外的知识",
+        ],
+    ),
+    "answer_relevancy": MetricRule(
+        warning_threshold=0.7,
+        critical_threshold=0.5,
+        higher_is_better=True,
+        root_causes=[
+            "回答偏离问题主旨或包含大量冗余内容",
+            "查询改写后问题语义漂移",
+            "生成 prompt 格式约束不足",
+        ],
+        suggested_actions=[
+            "优化查询改写 prompt，确保改写后语义不偏移",
+            "在生成 prompt 中加入'简洁准确、直接回答问题'的约束",
+            "检查低分样本的回答是否存在格式冗余或话题偏移",
+        ],
+    ),
+    "context_recall": MetricRule(
+        warning_threshold=0.7,
+        critical_threshold=0.5,
+        higher_is_better=True,
+        root_causes=[
+            "检索未能召回标准答案所涉及的关键信息",
+            "单一查询未能覆盖问题的多个角度",
+            "过召回数量不足，关键片段被截断",
+        ],
+        suggested_actions=[
+            "启用多查询扩展（use_multi_query）覆盖不同措辞",
+            "对多跳问题启用问题分解（sub_questions）",
+            "加大过召回宽度（recall_top_k）",
+            "对颗粒度细的问题尝试 Step-back 双路检索",
+        ],
+    ),
+    "context_precision": MetricRule(
+        warning_threshold=0.6,
+        critical_threshold=0.4,
+        higher_is_better=True,
+        root_causes=[
+            "检索引入过多与问题无关的片段",
+            "重排未能将相关片段排在前列",
+            "缺少相关性过滤，噪声片段进入上下文",
+        ],
+        suggested_actions=[
+            "启用或优化 listwise 重排，将相关片段排在前列",
+            "启用上下文压缩（compression）过滤无关句子",
+            "启用相关性过滤（relevance_filter）丢弃明确无关片段",
+            "缩小 rerank_keep_k（如从 8 降到 5）",
+        ],
+    ),
+    "noise_sensitivity": MetricRule(
+        warning_threshold=0.3,   # higher is worse; trigger when mean > threshold
+        critical_threshold=0.5,
+        higher_is_better=False,
+        root_causes=[
+            "回答中包含检索到的噪声片段所引入的错误陈述",
+            "相关性过滤未能拦截干扰性片段",
+            "生成阶段对噪声片段未加区分地引用",
+        ],
+        suggested_actions=[
+            "启用相关性过滤（relevance_filter）拦截噪声",
+            "优化重排，将不相关片段排到截断点之后",
+            "在生成 prompt 中强调'来源冲突时并列陈述，不擅自下定论'",
+        ],
+    ),
+    "factual_correctness": MetricRule(
+        warning_threshold=0.6,
+        critical_threshold=0.4,
+        higher_is_better=True,
+        root_causes=[
+            "回答的事实陈述与标准答案存在偏差",
+            "检索未能命中标准答案所依据的关键片段",
+            "生成阶段对多个来源综合时产生事实错误",
+        ],
+        suggested_actions=[
+            "重点检查低分样本，确认是检索遗漏还是生成错误",
+            "提升 context_recall 以确保关键信息被检索到",
+            "对事实型问题将 temperature 降至 0",
+        ],
+    ),
+    "semantic_similarity": MetricRule(
+        warning_threshold=0.7,
+        critical_threshold=0.5,
+        higher_is_better=True,
+        root_causes=[
+            "回答语义与标准答案差距较大",
+            "回答过于简短或过于冗长，语义偏移",
+            "检索到的片段质量不足，导致生成内容偏离",
+        ],
+        suggested_actions=[
+            "检查低分样本的回答与标准答案的表述差异",
+            "优化生成 prompt 使回答更贴近标准表述风格",
+            "提升检索质量（context_recall / context_precision）",
+        ],
+    ),
+}
+
+
+@dataclass
+class Diagnosis:
+    """Diagnostic result for one metric that triggered a threshold."""
+    metric: str
+    mean_score: float
+    threshold: float          # the triggered threshold
+    severity: str             # "warning" | "critical"
+    root_causes: list[str] = field(default_factory=list)
+    suggested_actions: list[str] = field(default_factory=list)
+    low_samples: list[dict[str, Any]] = field(default_factory=list)
+
+
+def _mean_ignoring_nan(values: list[float]) -> float | None:
+    valid = [v for v in values if not math.isnan(v)]
+    if not valid:
+        return None
+    return sum(valid) / len(valid)
+
+
+def _select_low_samples(
+    rows: list[dict[str, Any]],
+    metric: str,
+    top_n: int,
+    higher_is_better: bool,
+) -> list[dict[str, Any]]:
+    """Return the top_n worst-scoring rows for a metric, excluding NaN."""
+    valid = [r for r in rows if metric in r and not math.isnan(float(r[metric]))]
+    sorted_rows = sorted(valid, key=lambda r: float(r[metric]), reverse=not higher_is_better)
+    worst = sorted_rows[:top_n]
+    keep_keys = {"sample_id", "question", "answer", "ground_truth", metric}
+    return [{k: v for k, v in row.items() if k in keep_keys} for row in worst]
+
+
+def diagnose(
+    score_rows: list[dict[str, Any]],
+    metrics: list[str],
+    top_low_samples: int = 3,
+) -> list[Diagnosis]:
+    """Analyse score_rows and return a Diagnosis for each metric below threshold.
+
+    Args:
+        score_rows: List of per-sample score dicts (from EvaluationResult.score_rows).
+        metrics: Metric names to evaluate (from Scenario.metrics).
+        top_low_samples: How many worst-scoring samples to attach per diagnosis.
+
+    Returns:
+        List of Diagnosis objects, one per triggered metric. Empty if all OK.
+    """
+    diagnoses: list[Diagnosis] = []
+
+    for metric in metrics:
+        rule = METRIC_RULES.get(metric)
+        if rule is None:
+            continue  # unknown metric, skip
+
+        values = []
+        for row in score_rows:
+            raw = row.get(metric)
+            if raw is None:
+                continue
+            try:
+                v = float(raw)
+            except (TypeError, ValueError):
+                continue
+            values.append(v)
+
+        if not values:
+            continue
+
+        mean = _mean_ignoring_nan(values)
+        if mean is None:
+            continue
+
+        # Determine severity (direction-aware)
+        if rule.higher_is_better:
+            if mean < rule.critical_threshold:
+                severity = "critical"
+                threshold = rule.critical_threshold
+            elif mean < rule.warning_threshold:
+                severity = "warning"
+                threshold = rule.warning_threshold
+            else:
+                continue  # above warning threshold → no diagnosis
+        else:
+            # lower is better (noise_sensitivity)
+            if mean > rule.critical_threshold:
+                severity = "critical"
+                threshold = rule.critical_threshold
+            elif mean > rule.warning_threshold:
+                severity = "warning"
+                threshold = rule.warning_threshold
+            else:
+                continue
+
+        low_samples = _select_low_samples(score_rows, metric, top_low_samples, rule.higher_is_better)
+
+        diagnoses.append(Diagnosis(
+            metric=metric,
+            mean_score=round(mean, 4),
+            threshold=threshold,
+            severity=severity,
+            root_causes=list(rule.root_causes),
+            suggested_actions=list(rule.suggested_actions),
+            low_samples=low_samples,
+        ))
+
+    return diagnoses

rag_eval/advisor/writer.py

@@ -0,0 +1,82 @@
+"""Write optimization advice to markdown file and emit log summary."""
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+
+from .rules import Diagnosis
+
+logger = logging.getLogger("rag_eval.advisor")
+
+
+def _format_log_summary(diagnoses: list[Diagnosis], advice_path: Path) -> str:
+    """Return a single-line log summary of triggered diagnoses."""
+    if not diagnoses:
+        return "[advisor] 所有指标正常，无需优化建议。"
+    parts = [f"{d.metric}({d.mean_score:.2f}, {d.severity})" for d in diagnoses]
+    triggered = " ".join(parts)
+    return f"[advisor] 触发诊断 {len(diagnoses)} 项: {triggered}  →  {advice_path}"
+
+
+def _build_fallback_report(diagnoses: list[Diagnosis]) -> str:
+    """Build a rules-only report when LLM analysis is unavailable."""
+    if not diagnoses:
+        return ""
+    lines = ["## 规则诊断（LLM 分析不可用）\n"]
+    for d in diagnoses:
+        lines.append(f"### {d.metric}  [{d.severity}]  均值={d.mean_score:.4f}")
+        lines.append("\n**可能原因：**")
+        for cause in d.root_causes:
+            lines.append(f"- {cause}")
+        lines.append("\n**建议动作：**")
+        for action in d.suggested_actions:
+            lines.append(f"- {action}")
+        lines.append("")
+    return "\n".join(lines)
+
+
+def write_advice(
+    diagnoses: list[Diagnosis],
+    llm_markdown: str,
+    advice_path: Path,
+    scenario_name: str,
+    run_id: str,
+    judge_model: str,
+) -> None:
+    """Write optimization_advice.md and emit a log summary line.
+
+    Args:
+        diagnoses: List of Diagnosis from rules.diagnose().
+        llm_markdown: LLM-generated Markdown body. Empty string triggers fallback.
+        advice_path: Full path to write the .md file.
+        scenario_name: Human-readable scenario identifier for the report header.
+        run_id: Run identifier string.
+        judge_model: Model used for LLM analysis (shown in header).
+    """
+    advice_path.parent.mkdir(parents=True, exist_ok=True)
+
+    from rag_eval.shared.utils import utc_now_iso
+    header_lines = [
+        f"# 优化建议报告 — {scenario_name}",
+        "",
+        f"- run_id: `{run_id}`",
+        f"- 生成时间: `{utc_now_iso()}`",
+        f"- judge_model: `{judge_model}`",
+        "",
+        "---",
+        "",
+    ]
+
+    if not diagnoses:
+        body = "## ✅ 未发现明显指标异常\n\n所有指标均在正常范围内，当前 RAG 链路表现良好。\n"
+    elif llm_markdown:
+        body = llm_markdown
+    else:
+        body = _build_fallback_report(diagnoses)
+
+    content = "\n".join(header_lines) + body
+    advice_path.write_text(content, encoding="utf-8")
+
+    summary = _format_log_summary(diagnoses, advice_path)
+    logger.info(summary)
+    logger.info("[advisor] 优化建议已写出: %s", advice_path)

rag_eval/config/loader.py

@@ -61,6 +61,7 @@ def load_scenario(path: str | Path) -> Scenario:
             max_samples=model.runtime.max_samples,
         ),
         source_path=scenario_path,
+        optimization_advisor=model.optimization_advisor,
     )
     # Run cross-field checks after all relative paths have been resolved.
     validate_scenario(scenario)

rag_eval/config/schema.py

@@ -54,6 +54,7 @@ class ScenarioModel(BaseModel):
     metrics: list[str]
     output_dir: str
     runtime: RuntimeConfigModel = Field(default_factory=RuntimeConfigModel)
+    optimization_advisor: bool = False
 
     @field_validator("metrics")
     @classmethod

rag_eval/execution/evaluator.py

@@ -3,6 +3,8 @@
 from __future__ import annotations
 
 import asyncio
+import logging
+import time
 from typing import Any
 
 from rag_eval.adapters.base import AppAdapter
@@ -13,6 +15,8 @@ from rag_eval.metrics.pipeline import MetricPipeline
 from rag_eval.shared.models import EvaluationResult, InvalidSample, NormalizedSample, Scenario
 from rag_eval.shared.utils import utc_now_iso
 
+logger = logging.getLogger("rag_eval.execution.evaluator")
+
 
 class Evaluator:
     """Coordinate dataset loading, optional app execution, and metric scoring."""
@@ -31,27 +35,61 @@ class Evaluator:
     def evaluate(self) -> EvaluationResult:
         """Execute the full evaluation flow and return the collected results."""
         started_at = utc_now_iso()
+        scenario_name = self.scenario.scenario_name
+        mode = self.scenario.mode
+        logger.info("=" * 60)
+        logger.info("[eval] START  scenario=%s  mode=%s", scenario_name, mode)
+        logger.info("[eval] dataset=%s", self.scenario.dataset.path)
+        logger.info("[eval] metrics=%s", list(self.scenario.metrics))
+        logger.info("[eval] judge=%s  embed=%s", self.scenario.judge_model, self.scenario.embedding_model)
+
         raw_records = load_dataset_records(self.scenario.dataset.path)
+        logger.info("[eval] raw_records=%d", len(raw_records))
+
         samples, invalid_samples = normalize_records(
             raw_records,
             mode=self.scenario.mode,
             max_samples=self.scenario.runtime.max_samples,
         )
+        logger.info("[eval] normalized: valid=%d  invalid=%d", len(samples), len(invalid_samples))
 
         if self.scenario.mode == "online":
-            # Online mode enriches each sample by calling the target application first.
+            logger.info("[eval] online mode: calling app adapter for %d samples ...", len(samples))
+            t0 = time.monotonic()
             samples, online_invalids = asyncio.run(self._enrich_online_samples(samples))
+            elapsed = time.monotonic() - t0
             invalid_samples.extend(online_invalids)
+            logger.info(
+                "[eval] adapter done: enriched=%d  adapter_invalids=%d  elapsed=%.1fs",
+                len(samples), len(online_invalids), elapsed,
+            )
 
+        logger.info("[eval] scoring %d samples with metric pipeline ...", len(samples))
+        t0 = time.monotonic()
         metric_scores = asyncio.run(
             self.metric_pipeline.score_samples(
                 samples,
                 max_concurrency=self.scenario.runtime.metric_limit(),
             )
         )
+        elapsed = time.monotonic() - t0
+        logger.info("[eval] metric scoring done  elapsed=%.1fs", elapsed)
+
         finished_at = utc_now_iso()
         score_rows = [self._merge_score(sample, score) for sample, score in zip(samples, metric_scores)]
+
+        # Summary of NaN rates per metric
+        import math
+        for metric_name in self.scenario.metrics:
+            nan_count = sum(1 for row in score_rows if math.isnan(float(row.get(metric_name, float("nan")) or float("nan"))))
+            logger.info("[eval] %-22s  NaN=%d/%d (%.0f%%)",
+                        metric_name, nan_count, len(score_rows),
+                        100 * nan_count / len(score_rows) if score_rows else 0)
+
         run_id = finished_at.replace(":", "-")
+        logger.info("[eval] DONE  run_id=%s  total_valid=%d  total_invalid=%d",
+                    run_id, len(samples), len(invalid_samples))
+        logger.info("=" * 60)
         return EvaluationResult(
             scenario=self.scenario,
             run_id=run_id,
@@ -72,13 +110,27 @@ class Evaluator:
 
         valid: list[NormalizedSample] = []
         invalid: list[InvalidSample] = []
+        total = len(samples)
 
-        async def enrich_with_capture(sample: NormalizedSample) -> NormalizedSample | InvalidSample:
+        async def enrich_with_capture(idx: int, sample: NormalizedSample) -> NormalizedSample | InvalidSample:
             """Convert adapter exceptions into invalid samples instead of aborting the run."""
+            sid = sample.sample_id[:12]
+            logger.debug("[adapter] [%d/%d] calling adapter  sample=%s  question=%r",
+                         idx + 1, total, sid, (sample.question or "")[:60])
+            t0 = time.monotonic()
             try:
-                return await self.app_adapter.enrich_sample(sample)
+                result = await self.app_adapter.enrich_sample(sample)
+                elapsed = time.monotonic() - t0
+                ans_len = len(result.answer or "")
+                ctx_count = len(result.contexts or [])
+                logger.info("[adapter] [%d/%d] OK  sample=%-12s  ans_len=%d  ctx_count=%d  elapsed=%.1fs",
+                            idx + 1, total, sid, ans_len, ctx_count, elapsed)
+                return result
             except Exception as exc:
+                elapsed = time.monotonic() - t0
                 error_type = type(exc).__name__
+                logger.warning("[adapter] [%d/%d] FAIL  sample=%-12s  %s: %s  (elapsed=%.1fs)",
+                               idx + 1, total, sid, error_type, exc, elapsed)
                 return InvalidSample(
                     sample_id=sample.sample_id,
                     error=f"adapter failed [{error_type}]: {exc}",
@@ -86,8 +138,8 @@ class Evaluator:
                 )
 
         factories = [
-            (lambda sample=sample: enrich_with_capture(sample))
+            (lambda _idx=i, _sample=sample: enrich_with_capture(_idx, _sample))
-            for sample in samples
+            for i, sample in enumerate(samples)
         ]
         results = await gather_with_limit(factories, self.scenario.runtime.app_limit())
 
@@ -102,6 +154,8 @@ class Evaluator:
             if not sample.contexts:
                 errors.append("adapter returned empty contexts")
             if errors:
+                logger.warning("[adapter] incomplete payload  sample=%s  errors=%s",
+                               sample.sample_id[:12], errors)
                 invalid.append(
                     InvalidSample(
                         sample_id=sample.sample_id,
@@ -111,6 +165,9 @@ class Evaluator:
                 )
                 continue
             valid.append(sample)
+
+        logger.info("[adapter] enrichment summary: valid=%d  invalid=%d  of total=%d",
+                    len(valid), len(invalid), total)
         return valid, invalid
 
     def _merge_score(self, sample: NormalizedSample, score: Any) -> dict[str, Any]:

rag_eval/execution/runner.py

@@ -2,16 +2,42 @@
 
 from __future__ import annotations
 
+import logging
+import sys
+from pathlib import Path
+
 from rag_eval.adapters.http import HttpAppAdapter
 from rag_eval.adapters.python import PythonFunctionAdapter
+from rag_eval.advisor import run_advisor
 from rag_eval.config.loader import load_scenario
-from rag_eval.metrics.factory import build_metric_pipeline
+from rag_eval.metrics.factory import build_models, build_metric_pipeline
 from rag_eval.reporting.writers import write_run_artifacts
 from rag_eval.settings import EvaluationSettings
 from rag_eval.shared.models import Scenario
 
 from .evaluator import Evaluator
 
+logger = logging.getLogger("rag_eval.execution.runner")
+
+
+def _setup_logging(log_file: Path | None = None, level: int = logging.INFO) -> None:
+    """Configure root logger: always write to stderr, optionally also to a file."""
+    fmt = "%(asctime)s  %(levelname)-8s  %(name)s  %(message)s"
+    datefmt = "%H:%M:%S"
+
+    handlers: list[logging.Handler] = [logging.StreamHandler(sys.stderr)]
+    if log_file is not None:
+        log_file.parent.mkdir(parents=True, exist_ok=True)
+        fh = logging.FileHandler(log_file, encoding="utf-8")
+        fh.setFormatter(logging.Formatter(fmt, datefmt=datefmt))
+        handlers.append(fh)
+
+    logging.basicConfig(level=level, format=fmt, datefmt=datefmt, handlers=handlers, force=True)
+    # Also show ragas internal logs at WARNING so we can see LLM errors
+    logging.getLogger("ragas").setLevel(logging.WARNING)
+    logging.getLogger("httpx").setLevel(logging.WARNING)
+    logging.getLogger("openai").setLevel(logging.WARNING)
+
 
 def build_adapter(scenario: Scenario):
     """Instantiate the adapter required by the resolved scenario, if any."""
@@ -27,16 +53,32 @@ def build_adapter(scenario: Scenario):
 def run_scenario(
     scenario_path: str,
     settings: EvaluationSettings | None = None,
+    log_file: Path | None = None,
+    log_level: int = logging.INFO,
 ):
     """Run one scenario end to end and persist its reporting artifacts."""
+    _setup_logging(log_file=log_file, level=log_level)
+    logger.info("[runner] run_scenario  path=%s", scenario_path)
+
     settings = settings or EvaluationSettings()
     if not settings.openai_api_key:
         raise EnvironmentError("OPENAI_API_KEY must be set before running the evaluator.")
 
     scenario = load_scenario(scenario_path)
+    logger.info("[runner] scenario loaded: name=%s  mode=%s  max_samples=%s",
+                scenario.scenario_name, scenario.mode, scenario.runtime.max_samples)
+
+    # Build models once; reuse llm in both MetricPipeline and advisor.
+    llm, embeddings = build_models(scenario.judge_model, scenario.embedding_model, settings)
+
     adapter = build_adapter(scenario)
-    pipeline = build_metric_pipeline(scenario, settings)
+    pipeline = build_metric_pipeline(scenario, settings, llm=llm, embeddings=embeddings)
     evaluator = Evaluator(scenario=scenario, metric_pipeline=pipeline, app_adapter=adapter)
     result = evaluator.evaluate()
     write_run_artifacts(result)
+    logger.info("[runner] artifacts written for run_id=%s", result.run_id)
+
+    # Optimization advisor — runs only if scenario.optimization_advisor is True.
+    run_advisor(result, scenario, llm)
+
     return result

rag_eval/metrics/factory.py

@@ -18,7 +18,10 @@ from ragas.metrics.collections import (
     AnswerRelevancy,
     ContextPrecision,
     ContextRecall,
+    FactualCorrectness,
     Faithfulness,
+    NoiseSensitivity,
+    SemanticSimilarity,
 )
 
 from .pipeline import MetricPipeline
@@ -39,19 +42,34 @@ def build_models(
 def build_metric_pipeline(
     scenario: Scenario,
     settings: EvaluationSettings,
+    llm: Any | None = None,
+    embeddings: Any | None = None,
 ) -> MetricPipeline:
-    """Build a metric pipeline containing only the metrics requested by the scenario."""
+    """Build a metric pipeline containing only the metrics requested by the scenario.
+
+    If llm and embeddings are provided (pre-built by the caller), they are reused.
+    Otherwise, new instances are created from scenario + settings.
+    """
+    if llm is None or embeddings is None:
         llm, embeddings = build_models(
             scenario.judge_model,
             scenario.embedding_model,
             settings,
         )
+
     # Build the full registry once, then slice it by configured metric names.
     registry: dict[str, Any] = {
         "faithfulness": Faithfulness(llm=llm),
         "answer_relevancy": AnswerRelevancy(llm=llm, embeddings=embeddings),
         "context_recall": ContextRecall(llm=llm),
         "context_precision": ContextPrecision(llm=llm),
+        # Robustness / end-to-end metrics (架构设计 §10.2).
+        # NoiseSensitivity mode='relevant': sensitivity to noise from relevant contexts.
+        "noise_sensitivity": NoiseSensitivity(llm=llm),
+        # FactualCorrectness mode='f1': balances claim precision and recall vs. ground truth.
+        "factual_correctness": FactualCorrectness(llm=llm),
+        # SemanticSimilarity: embedding cosine between answer and ground truth (no LLM call).
+        "semantic_similarity": SemanticSimilarity(embeddings=embeddings),
     }
     return MetricPipeline(
         metrics={name: registry[name] for name in scenario.metrics},

rag_eval/metrics/pipeline.py

@@ -3,12 +3,16 @@
 from __future__ import annotations
 
 import asyncio
+import logging
 import math
+import time
 from dataclasses import dataclass
 from typing import Any
 
 from rag_eval.shared.models import MetricScore, NormalizedSample
 
+logger = logging.getLogger("rag_eval.metrics.pipeline")
+
 
 @dataclass(slots=True)
 class MetricPipeline:
@@ -22,12 +26,43 @@ class MetricPipeline:
         results = {name: math.nan for name in self.metrics}
         errors: list[str] = []
 
+        sid = sample.sample_id[:12]
+        ans_len = len(sample.answer or "")
+        ctx_count = len(sample.contexts or [])
+        logger.debug(
+            "[score] sample=%s  ans_len=%d  ctx_count=%d  question=%r",
+            sid, ans_len, ctx_count,
+            (sample.question or "")[:80],
+        )
+
         for name, metric in self.metrics.items():
+            t0 = time.monotonic()
             try:
                 result = await self._run_metric(name, metric, sample)
-                results[name] = float(result.value)
+                score_val = float(result.value)
+                results[name] = score_val
+                elapsed = time.monotonic() - t0
+                logger.info(
+                    "[metric OK ] sample=%-12s  %-20s  score=%.4f  elapsed=%.1fs",
+                    sid, name, score_val, elapsed,
+                )
+            except asyncio.TimeoutError:
+                elapsed = time.monotonic() - t0
+                msg = f"timeout after {self.metric_timeout_seconds}s"
+                errors.append(f"{name}: {msg}")
+                logger.warning(
+                    "[metric TMO] sample=%-12s  %-20s  TIMEOUT after %.1fs",
+                    sid, name, elapsed,
+                )
             except Exception as exc:
+                elapsed = time.monotonic() - t0
+                exc_type = type(exc).__name__
                 errors.append(f"{name}: {exc}")
+                logger.warning(
+                    "[metric ERR] sample=%-12s  %-20s  %s: %s  (elapsed=%.1fs)",
+                    sid, name, exc_type, exc, elapsed,
+                )
+
         return MetricScore(metrics=results, error=" | ".join(errors))
 
     async def _run_metric(self, name: str, metric: Any, sample: NormalizedSample) -> Any:
@@ -59,6 +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}")
 
@@ -72,11 +124,22 @@ class MetricPipeline:
         max_concurrency: int,
     ) -> list[MetricScore]:
         """Score all samples while respecting the configured concurrency limit."""
+        total = len(samples)
+        logger.info("[pipeline] scoring %d samples  concurrency=%d  timeout=%ss",
+                    total, max_concurrency, self.metric_timeout_seconds)
         semaphore = asyncio.Semaphore(max(1, max_concurrency))
+        completed = 0
 
-        async def guarded(sample: NormalizedSample) -> MetricScore:
+        async def guarded(idx: int, sample: NormalizedSample) -> MetricScore:
             """Throttle a single sample-scoring coroutine with the shared semaphore."""
+            nonlocal completed
             async with semaphore:
-                return await self.score_sample(sample)
+                result = await self.score_sample(sample)
+                completed += 1
+                nan_metrics = [k for k, v in result.metrics.items() if math.isnan(v)]
+                status = f"NaN={nan_metrics}" if nan_metrics else "all OK"
+                logger.info("[pipeline] progress %d/%d  sample=%-12s  %s",
+                            completed, total, sample.sample_id[:12], status)
+                return result
 
-        return await asyncio.gather(*(guarded(sample) for sample in samples))
+        return await asyncio.gather(*(guarded(i, s) for i, s in enumerate(samples)))

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）
 }

rag_eval/reporting/artifacts.py

@@ -17,4 +17,5 @@ def build_artifact_paths(output_dir: Path, run_id: str) -> RunArtifactPaths:
         invalid_csv=run_dir / "invalid.csv",
         summary_md=run_dir / "summary.md",
         metadata_json=run_dir / "metadata.json",
+        advice_md=run_dir / "optimization_advice.md",
     )

rag_eval/shared/models.py

@@ -76,6 +76,7 @@ class Scenario:
     runtime: RuntimeConfig = field(default_factory=RuntimeConfig)
     app_adapter: AppAdapterConfig | None = None
     source_path: Path | None = None
+    optimization_advisor: bool = False
 
     def snapshot(self) -> dict[str, Any]:
         """Serialize the scenario into a reporting-friendly dictionary snapshot."""
@@ -159,3 +160,4 @@ class RunArtifactPaths:
     invalid_csv: Path
     summary_md: Path
     metadata_json: Path
+    advice_md: Path | None = None

run_eval.bat

@@ -0,0 +1,107 @@
+@echo off
+setlocal enabledelayedexpansion
+
+:: ============================================================
+::  run_eval.bat  -  Run a RAGAS evaluation scenario with logs
+::
+::  Usage:
+::    run_eval.bat                          (uses default online scenario)
+::    run_eval.bat offline                  (runs offline smoke scenario)
+::    run_eval.bat path\to\scenario.yaml    (any custom scenario)
+::    run_eval.bat offline DEBUG            (second arg = log level)
+:: ============================================================
+
+cd /d "%~dp0"
+
+echo.
+echo ============================================================
+echo   Siemens RAGAS  -  Evaluation Runner
+echo ============================================================
+echo.
+
+:: ----------------------------------------------------------------
+:: 1. Resolve scenario path  (arg1)
+:: ----------------------------------------------------------------
+set "SCENARIO=%~1"
+if "%SCENARIO%"=="" set "SCENARIO=online"
+
+if /i "%SCENARIO%"=="online" (
+    set "SCENARIO=scenarios\online\siemens-pdf-question-bank-online.yaml"
+)
+if /i "%SCENARIO%"=="offline" (
+    set "SCENARIO=scenarios\offline\siemens-pdf-offline-smoke.yaml"
+)
+
+if not exist "%SCENARIO%" (
+    echo [ERROR] Scenario file not found: %SCENARIO%
+    echo.
+    echo Usage examples:
+    echo   run_eval.bat                    - online eval (default)
+    echo   run_eval.bat offline            - offline smoke
+    echo   run_eval.bat path\to\file.yaml  - custom scenario
+    goto :error
+)
+echo [OK] Scenario : %SCENARIO%
+
+:: ----------------------------------------------------------------
+:: 2. Resolve log level  (arg2, default INFO)
+:: ----------------------------------------------------------------
+set "LOG_LEVEL=%~2"
+if "%LOG_LEVEL%"=="" set "LOG_LEVEL=INFO"
+echo [OK] Log level: %LOG_LEVEL%
+
+:: ----------------------------------------------------------------
+:: 3. Create logs dir and build timestamped log filename
+:: ----------------------------------------------------------------
+if not exist "logs" mkdir logs
+for /f "tokens=1-3 delims=/-" %%a in ("%DATE%") do (
+    set "YMD=%%c-%%a-%%b"
+)
+for /f "tokens=1-3 delims=:." %%a in ("%TIME: =0%") do (
+    set "HMS=%%a%%b%%c"
+)
+set "LOG_FILE=logs\eval_%YMD%_%HMS%.log"
+echo [OK] Log file : %LOG_FILE%
+
+echo.
+echo ============================================================
+echo   Starting evaluation...
+echo   (Logs also written to %LOG_FILE%)
+echo   Press Ctrl+C to abort
+echo ============================================================
+echo.
+
+:: ----------------------------------------------------------------
+:: 4. Run evaluation with UTF-8 and logging
+:: ----------------------------------------------------------------
+set PYTHONIOENCODING=utf-8
+set PYTHONPATH=.
+
+python main.py ^
+    --scenario "%SCENARIO%" ^
+    --log-file "%LOG_FILE%" ^
+    --log-level %LOG_LEVEL%
+
+if errorlevel 1 (
+    echo.
+    echo [ERROR] Evaluation failed. Check log: %LOG_FILE%
+    goto :error
+)
+
+echo.
+echo ============================================================
+echo   Evaluation complete!
+echo   Log saved to: %LOG_FILE%
+echo   Open the web console to view results: start.bat
+echo ============================================================
+echo.
+pause
+exit /b 0
+
+:error
+echo.
+echo ============================================================
+echo   Evaluation failed. See error above or check log file.
+echo ============================================================
+pause
+exit /b 1

run_eval.ps1

@@ -0,0 +1,96 @@
+# run_eval.ps1 - Siemens RAGAS Evaluation Runner
+# Usage:
+#   .\run_eval.ps1                         # online eval (default)
+#   .\run_eval.ps1 offline                 # offline smoke
+#   .\run_eval.ps1 path\to\scenario.yaml   # custom scenario
+#   .\run_eval.ps1 online DEBUG            # second arg = log level (DEBUG/INFO/WARNING)
+# Or: powershell -ExecutionPolicy Bypass -File run_eval.ps1 [scenario] [log-level]
+
+param(
+    [string]$Scenario = "online",
+    [string]$LogLevel = "INFO"
+)
+
+$ErrorActionPreference = "Stop"
+Set-Location $PSScriptRoot
+
+Write-Host ""
+Write-Host "============================================================" -ForegroundColor Cyan
+Write-Host "  Siemens RAGAS  -  Evaluation Runner" -ForegroundColor Cyan
+Write-Host "============================================================" -ForegroundColor Cyan
+Write-Host ""
+
+# ----------------------------------------------------------------
+# 1. Resolve scenario path
+# ----------------------------------------------------------------
+$scenarioMap = @{
+    "online"  = "scenarios\online\siemens-pdf-question-bank-online.yaml"
+    "offline" = "scenarios\offline\siemens-pdf-offline-smoke.yaml"
+}
+if ($scenarioMap.ContainsKey($Scenario.ToLower())) {
+    $Scenario = $scenarioMap[$Scenario.ToLower()]
+}
+if (-not (Test-Path $Scenario)) {
+    Write-Host "[ERROR] Scenario file not found: $Scenario" -ForegroundColor Red
+    Write-Host ""
+    Write-Host "Usage examples:"
+    Write-Host "  .\run_eval.ps1                    - online eval (default)"
+    Write-Host "  .\run_eval.ps1 offline            - offline smoke"
+    Write-Host "  .\run_eval.ps1 path\to\file.yaml  - custom scenario"
+    Read-Host "Press Enter to exit"
+    exit 1
+}
+Write-Host "[OK] Scenario : $Scenario" -ForegroundColor Green
+
+# ----------------------------------------------------------------
+# 2. Validate log level
+# ----------------------------------------------------------------
+$validLevels = @("DEBUG", "INFO", "WARNING", "ERROR")
+if ($validLevels -notcontains $LogLevel.ToUpper()) {
+    Write-Host "[WARN] Unknown log level '$LogLevel', defaulting to INFO" -ForegroundColor Yellow
+    $LogLevel = "INFO"
+}
+Write-Host "[OK] Log level: $LogLevel" -ForegroundColor Green
+
+# ----------------------------------------------------------------
+# 3. Create logs dir with timestamped filename
+# ----------------------------------------------------------------
+if (-not (Test-Path "logs")) { New-Item -ItemType Directory "logs" | Out-Null }
+$timestamp = Get-Date -Format "yyyy-MM-dd_HHmmss"
+$logFile = "logs\eval_$timestamp.log"
+Write-Host "[OK] Log file : $logFile" -ForegroundColor Green
+
+Write-Host ""
+Write-Host "============================================================" -ForegroundColor Cyan
+Write-Host "  Starting evaluation..." -ForegroundColor Cyan
+Write-Host "  Logs also written to: $logFile" -ForegroundColor Cyan
+Write-Host "  Press Ctrl+C to abort" -ForegroundColor Yellow
+Write-Host "============================================================" -ForegroundColor Cyan
+Write-Host ""
+
+# ----------------------------------------------------------------
+# 4. Run evaluation
+# ----------------------------------------------------------------
+$env:PYTHONIOENCODING = "utf-8"
+$env:PYTHONPATH = "."
+
+& python main.py `
+    --scenario $Scenario `
+    --log-file $logFile `
+    --log-level $LogLevel.ToUpper()
+
+if ($LASTEXITCODE -ne 0) {
+    Write-Host ""
+    Write-Host "[ERROR] Evaluation failed. Check log: $logFile" -ForegroundColor Red
+    Read-Host "Press Enter to exit"
+    exit 1
+}
+
+Write-Host ""
+Write-Host "============================================================" -ForegroundColor Green
+Write-Host "  Evaluation complete!" -ForegroundColor Green
+Write-Host "  Log saved to: $logFile" -ForegroundColor Green
+Write-Host "  Open the web console to view results: start.bat" -ForegroundColor Cyan
+Write-Host "============================================================" -ForegroundColor Green
+Write-Host ""
+Read-Host "Press Enter to exit"

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

@@ -3,20 +3,24 @@ mode: online
 dataset: ../../datasets/raw/generated/siemens-pdf-question-bank.csv
 judge_model: deepseek-v4-flash
 embedding_model: text-embedding-v3
+optimization_advisor: true
 metrics:
 - faithfulness
 - answer_relevancy
 - context_recall
 - context_precision
+- noise_sensitivity
+- factual_correctness
+- semantic_similarity
 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

scripts/smoke_advisor.py

@@ -0,0 +1,59 @@
+"""Offline smoke-check for the advisor module wiring (no network required)."""
+import math
+import sys
+import tempfile
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from rag_eval.advisor.rules import diagnose
+from rag_eval.advisor.writer import write_advice, _format_log_summary
+
+# Simulate score_rows with low faithfulness and high noise_sensitivity
+rows = [
+    {
+        "sample_id": f"s{i}",
+        "question": f"问题{i}：西门子CT扫描的Flash技术原理是什么？",
+        "answer": f"答案{i}：Flash技术采用双源CT扫描",
+        "ground_truth": f"标准答案{i}：Flash扫描利用双源CT和大螺距实现超低辐射剂量扫描",
+        "faithfulness": 0.3 + i * 0.05,
+        "noise_sensitivity": 0.4 + i * 0.02,
+        "context_recall": 0.75,
+        "semantic_similarity": 0.65,
+    }
+    for i in range(5)
+]
+
+diags = diagnose(rows, metrics=["faithfulness", "noise_sensitivity", "context_recall", "semantic_similarity"])
+print(f"Diagnosed {len(diags)} metric(s):")
+for d in diags:
+    print(f"  {d.metric}: mean={d.mean_score}, severity={d.severity}, low_samples={len(d.low_samples)}")
+
+assert len(diags) >= 2, f"Expected at least 2 diagnoses, got {len(diags)}"
+metrics_hit = {d.metric for d in diags}
+assert "faithfulness" in metrics_hit, "faithfulness should be triggered"
+assert "noise_sensitivity" in metrics_hit, "noise_sensitivity should be triggered"
+
+with tempfile.TemporaryDirectory() as tmp:
+    path = Path(tmp) / "optimization_advice.md"
+    write_advice(
+        diagnoses=diags,
+        llm_markdown="",  # fallback mode (no LLM)
+        advice_path=path,
+        scenario_name="smoke-test-siemens",
+        run_id="2026-06-16T00-00-00",
+        judge_model="deepseek-v4-flash",
+    )
+    content = path.read_text(encoding="utf-8")
+    assert "smoke-test-siemens" in content, "scenario name missing from report"
+    assert "faithfulness" in content, "faithfulness missing from report"
+    assert "noise_sensitivity" in content, "noise_sensitivity missing from report"
+    print(f"\nAdvice file ({len(content)} chars) — assertions OK")
+
+# Verify log summary format
+summary = _format_log_summary(diags, Path("optimization_advice.md"))
+print(f"\nLog summary length: {len(summary)} chars, faithfulness present: {'faithfulness' in summary}")
+assert "触发诊断" in summary
+assert "faithfulness" in summary
+
+print("\nSmoke check PASSED")

start.bat

@@ -56,7 +56,17 @@ if errorlevel 1 (
 )
 
 :: ----------------------------------------------------------------
-:: 4. Seed demo data if no runs exist yet
+:: 4. Ensure configs/ directory exists for LLM profile storage
+:: ----------------------------------------------------------------
+if not exist "configs" (
+    mkdir configs
+    echo [OK] Created configs/ directory for LLM profile storage.
+) else (
+    echo [OK] configs/ directory ready.
+)
+
+:: ----------------------------------------------------------------
+:: 5. Seed demo data if no runs exist yet
 :: ----------------------------------------------------------------
 if not exist "outputs\kba-knowledge-base-offline-baseline" (
     echo [INFO] No run data found. Generating demo data...
@@ -71,7 +81,7 @@ if not exist "outputs\kba-knowledge-base-offline-baseline" (
 )
 
 :: ----------------------------------------------------------------
-:: 5. Pick an available port
+:: 6. Pick an available port
 :: ----------------------------------------------------------------
 set PORT=8800
 netstat -ano 2>nul | findstr ":8800" | findstr "LISTENING" >nul 2>&1

start.ps1

@@ -58,7 +58,17 @@ if ($LASTEXITCODE -ne 0) {
 }
 
 # ----------------------------------------------------------------
-# 4. Seed demo data if missing
+# 4. Ensure configs/ directory exists for LLM profile storage
+# ----------------------------------------------------------------
+if (-not (Test-Path "configs")) {
+    New-Item -ItemType Directory "configs" | Out-Null
+    Write-Host "[OK] Created configs/ directory for LLM profile storage." -ForegroundColor Green
+} else {
+    Write-Host "[OK] configs/ directory ready." -ForegroundColor Green
+}
+
+# ----------------------------------------------------------------
+# 5. Seed demo data if missing
 # ----------------------------------------------------------------
 if (-not (Test-Path "outputs\kba-knowledge-base-offline-baseline")) {
     Write-Host "[INFO] No run data found. Generating demo data..." -ForegroundColor Yellow
@@ -73,7 +83,7 @@ if (-not (Test-Path "outputs\kba-knowledge-base-offline-baseline")) {
 }
 
 # ----------------------------------------------------------------
-# 5. Pick an available port
+# 6. Pick an available port
 # ----------------------------------------------------------------
 $PORT = 8800
 $inUse = netstat -ano 2>$null | Select-String ":$PORT\s" | Select-String "LISTENING"

tests/test_advisor_rules.py

@@ -0,0 +1,72 @@
+import math
+import unittest
+from rag_eval.advisor.rules import Diagnosis, diagnose, METRIC_RULES
+
+
+class TestDiagnosis(unittest.TestCase):
+    def _make_rows(self, metric: str, scores: list[float]) -> list[dict]:
+        return [{metric: s, "question": f"q{i}", "answer": f"a{i}",
+                 "ground_truth": f"gt{i}", "sample_id": f"s{i}"}
+                for i, s in enumerate(scores)]
+
+    def test_no_diagnosis_when_all_scores_above_threshold(self):
+        rows = self._make_rows("faithfulness", [0.8, 0.9, 0.85])
+        result = diagnose(rows, metrics=["faithfulness"])
+        self.assertEqual(result, [])
+
+    def test_warning_when_mean_below_warning_threshold(self):
+        rows = self._make_rows("faithfulness", [0.65, 0.62, 0.68])
+        result = diagnose(rows, metrics=["faithfulness"])
+        self.assertEqual(len(result), 1)
+        self.assertEqual(result[0].metric, "faithfulness")
+        self.assertEqual(result[0].severity, "warning")
+        self.assertAlmostEqual(result[0].mean_score, 0.65, places=2)
+
+    def test_critical_when_mean_below_critical_threshold(self):
+        rows = self._make_rows("faithfulness", [0.3, 0.4, 0.45])
+        result = diagnose(rows, metrics=["faithfulness"])
+        self.assertEqual(result[0].severity, "critical")
+
+    def test_low_samples_selected_are_bottom_three(self):
+        rows = self._make_rows("faithfulness", [0.1, 0.2, 0.3, 0.8, 0.9])
+        result = diagnose(rows, metrics=["faithfulness"])
+        self.assertEqual(len(result[0].low_samples), 3)
+        scores = [s["faithfulness"] for s in result[0].low_samples]
+        self.assertEqual(sorted(scores), [0.1, 0.2, 0.3])
+
+    def test_nan_scores_excluded_from_mean_and_low_samples(self):
+        rows = self._make_rows("faithfulness", [0.3, float("nan"), 0.4])
+        result = diagnose(rows, metrics=["faithfulness"])
+        self.assertEqual(len(result), 1)
+        for s in result[0].low_samples:
+            self.assertFalse(math.isnan(s["faithfulness"]))
+
+    def test_noise_sensitivity_direction_inverted(self):
+        # noise_sensitivity: higher is worse; threshold > 0.3 is warning
+        rows = self._make_rows("noise_sensitivity", [0.4, 0.45, 0.5])
+        result = diagnose(rows, metrics=["noise_sensitivity"])
+        self.assertEqual(len(result), 1)
+        self.assertEqual(result[0].metric, "noise_sensitivity")
+
+    def test_noise_sensitivity_no_diagnosis_when_low(self):
+        rows = self._make_rows("noise_sensitivity", [0.1, 0.15, 0.2])
+        result = diagnose(rows, metrics=["noise_sensitivity"])
+        self.assertEqual(result, [])
+
+    def test_skips_metric_not_in_rows(self):
+        rows = [{"faithfulness": 0.3, "question": "q", "answer": "a",
+                 "ground_truth": "gt", "sample_id": "s1"}]
+        result = diagnose(rows, metrics=["faithfulness", "context_recall"])
+        metrics_found = [d.metric for d in result]
+        self.assertIn("faithfulness", metrics_found)
+        self.assertNotIn("context_recall", metrics_found)
+
+    def test_all_seven_metrics_have_rules(self):
+        expected = {"faithfulness", "answer_relevancy", "context_recall",
+                    "context_precision", "noise_sensitivity",
+                    "factual_correctness", "semantic_similarity"}
+        self.assertEqual(set(METRIC_RULES.keys()), expected)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_advisor_writer.py

@@ -0,0 +1,113 @@
+import shutil
+import unittest
+from pathlib import Path
+
+from rag_eval.advisor.rules import Diagnosis
+from rag_eval.advisor.writer import write_advice, _format_log_summary
+
+
+class TestWriteAdvice(unittest.TestCase):
+    def setUp(self):
+        self.tmp = Path("tests/.tmp/test_advisor_writer")
+        shutil.rmtree(self.tmp, ignore_errors=True)
+        self.tmp.mkdir(parents=True, exist_ok=True)
+        self.advice_path = self.tmp / "optimization_advice.md"
+
+    def tearDown(self):
+        shutil.rmtree(self.tmp, ignore_errors=True)
+
+    def _make_diagnosis(self, metric="faithfulness", severity="warning"):
+        return Diagnosis(
+            metric=metric,
+            mean_score=0.55,
+            threshold=0.7,
+            severity=severity,
+            root_causes=["原因1", "原因2"],
+            suggested_actions=["建议1", "建议2"],
+            low_samples=[
+                {"sample_id": "s1", "question": "问题1", "answer": "答案1",
+                 "ground_truth": "标准1", metric: 0.4},
+            ],
+        )
+
+    def test_write_creates_file(self):
+        diag = self._make_diagnosis()
+        write_advice(
+            diagnoses=[diag],
+            llm_markdown="## faithfulness\n\nLLM 建议内容",
+            advice_path=self.advice_path,
+            scenario_name="test-scenario",
+            run_id="2026-01-01T00-00-00",
+            judge_model="deepseek-v4-flash",
+        )
+        self.assertTrue(self.advice_path.exists())
+
+    def test_write_contains_scenario_name_and_run_id(self):
+        diag = self._make_diagnosis()
+        write_advice(
+            diagnoses=[diag],
+            llm_markdown="## faithfulness\n\nLLM 建议",
+            advice_path=self.advice_path,
+            scenario_name="siemens-test",
+            run_id="2026-01-01T00-00-00",
+            judge_model="deepseek-v4-flash",
+        )
+        content = self.advice_path.read_text(encoding="utf-8")
+        self.assertIn("siemens-test", content)
+        self.assertIn("2026-01-01T00-00-00", content)
+
+    def test_write_contains_llm_markdown(self):
+        diag = self._make_diagnosis()
+        write_advice(
+            diagnoses=[diag],
+            llm_markdown="## faithfulness\n\n具体建议文本",
+            advice_path=self.advice_path,
+            scenario_name="test",
+            run_id="rid",
+            judge_model="model",
+        )
+        content = self.advice_path.read_text(encoding="utf-8")
+        self.assertIn("具体建议文本", content)
+
+    def test_write_fallback_when_no_llm_markdown(self):
+        """When llm_markdown is empty, writer emits rule-only report."""
+        diag = self._make_diagnosis()
+        write_advice(
+            diagnoses=[diag],
+            llm_markdown="",
+            advice_path=self.advice_path,
+            scenario_name="test",
+            run_id="rid",
+            judge_model="model",
+        )
+        content = self.advice_path.read_text(encoding="utf-8")
+        self.assertIn("faithfulness", content)
+        self.assertIn("原因1", content)
+
+    def test_log_summary_format(self):
+        diags = [
+            self._make_diagnosis("faithfulness", "critical"),
+            self._make_diagnosis("context_recall", "warning"),
+        ]
+        summary = _format_log_summary(diags, self.advice_path)
+        self.assertIn("faithfulness", summary)
+        self.assertIn("critical", summary)
+        self.assertIn("context_recall", summary)
+        self.assertIn("warning", summary)
+
+    def test_write_empty_diagnoses_still_creates_file(self):
+        write_advice(
+            diagnoses=[],
+            llm_markdown="",
+            advice_path=self.advice_path,
+            scenario_name="test",
+            run_id="rid",
+            judge_model="model",
+        )
+        self.assertTrue(self.advice_path.exists())
+        content = self.advice_path.read_text(encoding="utf-8")
+        self.assertIn("未发现明显指标异常", content)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/webapp/test_llm_profiles_api.py

@@ -0,0 +1,139 @@
+"""Integration tests for /api/llm-profiles endpoints."""
+import pytest
+from fastapi.testclient import TestClient
+
+
+@pytest.fixture()
+def client(tmp_path, monkeypatch):
+    """TestClient with a fresh ProfileManager backed by a temp file."""
+    store = tmp_path / "profiles.json"
+    import webapp.services.profile_manager as pm_mod
+    from webapp.services.profile_manager import ProfileManager
+    fresh_mgr = ProfileManager(store_path=store)
+    monkeypatch.setattr(pm_mod, "profile_manager", fresh_mgr)
+    import webapp.api.llm_profiles as api_mod
+    monkeypatch.setattr(api_mod, "profile_manager", fresh_mgr)
+
+    from webapp.server import create_app
+    return TestClient(create_app())
+
+
+def test_list_empty(client):
+    resp = client.get("/api/llm-profiles")
+    assert resp.status_code == 200
+    assert resp.json()["profiles"] == []
+
+
+def test_create_and_list(client):
+    body = {"name": "Test", "model": "m1", "base_url": "http://x/v1", "api_key": "k"}
+    resp = client.post("/api/llm-profiles", json=body)
+    assert resp.status_code == 201
+    data = resp.json()
+    assert data["name"] == "Test"
+    assert data["profile_id"] != ""
+
+    resp2 = client.get("/api/llm-profiles")
+    assert len(resp2.json()["profiles"]) == 1
+
+
+def test_update_profile(client):
+    body = {"name": "Old", "model": "m1", "base_url": "http://x/v1", "api_key": "k"}
+    pid = client.post("/api/llm-profiles", json=body).json()["profile_id"]
+
+    upd = {"name": "New", "model": "m2", "base_url": "http://x/v1", "api_key": "k", "timeout_seconds": 60}
+    resp = client.put(f"/api/llm-profiles/{pid}", json=upd)
+    assert resp.status_code == 200
+    assert resp.json()["name"] == "New"
+    assert resp.json()["timeout_seconds"] == 60
+
+
+def test_delete_profile(client):
+    body = {"name": "Del", "model": "m", "base_url": "http://x/v1", "api_key": "k"}
+    pid = client.post("/api/llm-profiles", json=body).json()["profile_id"]
+    resp = client.delete(f"/api/llm-profiles/{pid}")
+    assert resp.status_code == 200
+    assert resp.json()["deleted"] is True
+    assert len(client.get("/api/llm-profiles").json()["profiles"]) == 0
+
+
+def test_update_nonexistent(client):
+    resp = client.put("/api/llm-profiles/nope",
+                      json={"name": "X", "model": "m", "base_url": "http://x/v1", "api_key": "k"})
+    assert resp.status_code == 404
+
+
+def test_delete_nonexistent(client):
+    resp = client.delete("/api/llm-profiles/nope")
+    assert resp.status_code == 404
+
+
+# ---------------------------------------------------------------------------
+# YAML patcher tests
+# ---------------------------------------------------------------------------
+import yaml as yaml_lib
+from webapp.services.yaml_patcher import apply_profiles_to_scenario
+from webapp.models import LLMProfile
+
+
+def test_apply_judge_profile(tmp_path):
+    """Applying a judge profile patches judge_model in the YAML."""
+    scenario_file = tmp_path / "test-scenario.yaml"
+    scenario_file.write_text(
+        "scenario_name: test\nmode: offline\njudge_model: old-model\nembedding_model: emb\n"
+        "dataset: data.csv\nmetrics:\n- faithfulness\noutput_dir: outputs/test\n",
+        encoding="utf-8",
+    )
+    judge_p = LLMProfile(
+        profile_id="x", name="J", model="new-model",
+        base_url="http://x/v1", api_key="k", created_at="t", updated_at="t",
+    )
+    patched = apply_profiles_to_scenario(
+        scenario_path=str(scenario_file),
+        judge_profile=judge_p,
+        answer_profile=None,
+        dataset_profile=None,
+        _resolve_absolute=True,
+    )
+    assert "judge_model" in patched
+    data = yaml_lib.safe_load(scenario_file.read_text())
+    assert data["judge_model"] == "new-model"
+
+
+def test_apply_answer_profile(tmp_path):
+    """Applying an answer profile patches app_adapter.static_kwargs.model."""
+    scenario_file = tmp_path / "online.yaml"
+    scenario_file.write_text(
+        "scenario_name: online\nmode: online\njudge_model: j\nembedding_model: emb\n"
+        "dataset: d.csv\nmetrics:\n- faithfulness\noutput_dir: out\n"
+        "app_adapter:\n  type: python\n  callable: apps.foo:run\n"
+        "  static_kwargs:\n    model: old\n    source_chunks_path: chunks.jsonl\n",
+        encoding="utf-8",
+    )
+    answer_p = LLMProfile(
+        profile_id="y", name="A", model="new-answer-model",
+        base_url="http://x/v1", api_key="k", created_at="t", updated_at="t",
+    )
+    patched = apply_profiles_to_scenario(
+        scenario_path=str(scenario_file),
+        judge_profile=None,
+        answer_profile=answer_p,
+        dataset_profile=None,
+        _resolve_absolute=True,
+    )
+    assert "app_adapter.static_kwargs.model" in patched
+    data = yaml_lib.safe_load(scenario_file.read_text())
+    assert data["app_adapter"]["static_kwargs"]["model"] == "new-answer-model"
+
+
+def test_apply_no_profiles_returns_empty(tmp_path):
+    """When no profiles are given, no fields are patched."""
+    scenario_file = tmp_path / "noop.yaml"
+    scenario_file.write_text("scenario_name: noop\njudge_model: m\n", encoding="utf-8")
+    patched = apply_profiles_to_scenario(
+        scenario_path=str(scenario_file),
+        judge_profile=None,
+        answer_profile=None,
+        dataset_profile=None,
+        _resolve_absolute=True,
+    )
+    assert patched == []

tests/webapp/test_profile_manager.py

@@ -0,0 +1,100 @@
+import pytest
+from webapp.models import LLMProfile, ProfileApplyRequest, ProfileApplyResponse
+
+def test_llm_profile_defaults():
+    p = LLMProfile(
+        profile_id="abc",
+        name="Test",
+        model="gpt-4",
+        base_url="http://localhost/v1",
+        api_key="sk-test",
+    )
+    assert p.timeout_seconds == 30
+    assert p.created_at != ""
+    assert p.updated_at != ""
+
+def test_profile_apply_request_fields():
+    req = ProfileApplyRequest(
+        scenario_path="scenarios/offline/sample.yaml",
+        judge_profile_id="id1",
+        answer_profile_id="id2",
+        dataset_profile_id=None,
+    )
+    assert req.judge_profile_id == "id1"
+    assert req.dataset_profile_id is None
+
+def test_profile_apply_response():
+    resp = ProfileApplyResponse(scenario_path="scenarios/offline/sample.yaml", patched_fields=["judge_model"])
+    assert "judge_model" in resp.patched_fields
+
+
+# ---------------------------------------------------------------------------
+# ProfileManager service tests
+# ---------------------------------------------------------------------------
+import json
+from webapp.services.profile_manager import ProfileManager
+
+
+def _make_manager(tmp_path):
+    store = tmp_path / "profiles.json"
+    return ProfileManager(store_path=store)
+
+
+def test_create_profile(tmp_path):
+    mgr = _make_manager(tmp_path)
+    p = mgr.create(name="Local", model="deepseek-v4-flash",
+                   base_url="http://localhost/v1", api_key="sk-x")
+    assert p.profile_id != ""
+    assert p.name == "Local"
+
+
+def test_list_profiles(tmp_path):
+    mgr = _make_manager(tmp_path)
+    mgr.create(name="A", model="m1", base_url="http://a/v1", api_key="k1")
+    mgr.create(name="B", model="m2", base_url="http://b/v1", api_key="k2")
+    profiles = mgr.list_all()
+    assert len(profiles) == 2
+
+
+def test_get_profile(tmp_path):
+    mgr = _make_manager(tmp_path)
+    created = mgr.create(name="X", model="m", base_url="http://x/v1", api_key="k")
+    fetched = mgr.get(created.profile_id)
+    assert fetched is not None
+    assert fetched.name == "X"
+
+
+def test_update_profile(tmp_path):
+    mgr = _make_manager(tmp_path)
+    p = mgr.create(name="Old", model="m", base_url="http://x/v1", api_key="k")
+    updated = mgr.update(p.profile_id, name="New", model="m2",
+                         base_url="http://x/v1", api_key="k", timeout_seconds=60)
+    assert updated is not None
+    assert updated.name == "New"
+    assert updated.model == "m2"
+    assert updated.timeout_seconds == 60
+
+
+def test_delete_profile(tmp_path):
+    mgr = _make_manager(tmp_path)
+    p = mgr.create(name="Del", model="m", base_url="http://x/v1", api_key="k")
+    assert mgr.delete(p.profile_id) is True
+    assert mgr.get(p.profile_id) is None
+
+
+def test_persistence(tmp_path):
+    store = tmp_path / "profiles.json"
+    mgr1 = ProfileManager(store_path=store)
+    p = mgr1.create(name="Persist", model="m", base_url="http://x/v1", api_key="k")
+    mgr2 = ProfileManager(store_path=store)
+    assert mgr2.get(p.profile_id) is not None
+
+
+def test_get_nonexistent(tmp_path):
+    mgr = _make_manager(tmp_path)
+    assert mgr.get("does-not-exist") is None
+
+
+def test_delete_nonexistent(tmp_path):
+    mgr = _make_manager(tmp_path)
+    assert mgr.delete("does-not-exist") is False

webapp/api/llm_profiles.py

@@ -0,0 +1,96 @@
+"""CRUD routes for LLM profiles plus the scenario-patching apply endpoint."""
+
+from __future__ import annotations
+
+from fastapi import APIRouter, HTTPException
+
+from webapp.models import (
+    CreateProfileRequest,
+    LLMProfile,
+    ProfileApplyRequest,
+    ProfileApplyResponse,
+)
+from webapp.services.profile_manager import profile_manager
+from webapp.services.yaml_patcher import apply_profiles_to_scenario
+
+router = APIRouter(prefix="/api/llm-profiles", tags=["llm-profiles"])
+
+
+@router.get("", response_model=dict)
+def list_profiles() -> dict:
+    """Return all saved LLM profiles."""
+    return {"profiles": [p.model_dump() for p in profile_manager.list_all()]}
+
+
+@router.post("", status_code=201, response_model=LLMProfile)
+def create_profile(request: CreateProfileRequest) -> LLMProfile:
+    """Create a new LLM profile."""
+    return profile_manager.create(
+        name=request.name,
+        model=request.model,
+        base_url=request.base_url,
+        api_key=request.api_key,
+        timeout_seconds=request.timeout_seconds,
+    )
+
+
+@router.put("/{profile_id}", response_model=LLMProfile)
+def update_profile(profile_id: str, request: CreateProfileRequest) -> LLMProfile:
+    """Update an existing LLM profile by id."""
+    updated = profile_manager.update(
+        profile_id=profile_id,
+        name=request.name,
+        model=request.model,
+        base_url=request.base_url,
+        api_key=request.api_key,
+        timeout_seconds=request.timeout_seconds,
+    )
+    if updated is None:
+        raise HTTPException(status_code=404, detail=f"Profile not found: {profile_id}")
+    return updated
+
+
+@router.delete("/{profile_id}", response_model=dict)
+def delete_profile(profile_id: str) -> dict:
+    """Delete an LLM profile by id."""
+    deleted = profile_manager.delete(profile_id)
+    if not deleted:
+        raise HTTPException(status_code=404, detail=f"Profile not found: {profile_id}")
+    return {"deleted": True}
+
+
+@router.post("/apply", response_model=ProfileApplyResponse)
+def apply_profiles(request: ProfileApplyRequest) -> ProfileApplyResponse:
+    """Patch selected LLM profiles into the target scenario YAML file."""
+    role_profiles: dict[str, LLMProfile | None] = {
+        "judge": profile_manager.get(request.judge_profile_id) if request.judge_profile_id else None,
+        "answer": profile_manager.get(request.answer_profile_id) if request.answer_profile_id else None,
+        "dataset": profile_manager.get(request.dataset_profile_id) if request.dataset_profile_id else None,
+    }
+
+    missing = [
+        role
+        for role, pid in [
+            ("judge", request.judge_profile_id),
+            ("answer", request.answer_profile_id),
+            ("dataset", request.dataset_profile_id),
+        ]
+        if pid and role_profiles[role] is None
+    ]
+
+    if missing:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Profile(s) not found for roles: {', '.join(missing)}",
+        )
+
+    patched = apply_profiles_to_scenario(
+        scenario_path=request.scenario_path,
+        judge_profile=role_profiles["judge"],
+        answer_profile=role_profiles["answer"],
+        dataset_profile=role_profiles["dataset"],
+    )
+    return ProfileApplyResponse(
+        scenario_path=request.scenario_path,
+        patched_fields=patched,
+    )

webapp/models.py

@@ -2,11 +2,16 @@
 
 from __future__ import annotations
 
+from datetime import datetime, timezone
 from typing import Any
 
 from pydantic import BaseModel, Field
 
 
+def _utcnow_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
 class RunSummary(BaseModel):
     """Compact description of a single evaluation run for list views."""
 
@@ -68,6 +73,7 @@ class ReportData(BaseModel):
     groupings: dict[str, list[GroupStat]] = Field(default_factory=dict)
     lowest_samples: list[SampleScore] = Field(default_factory=list)
     summary_markdown: str = ""
+    advice_markdown: str = ""  # optimization_advice.md content (empty if not generated)
 
 
 class RunDetail(BaseModel):
@@ -114,6 +120,45 @@ class TriggerEvaluationResponse(BaseModel):
     task_id: str
 
 
+class LLMProfile(BaseModel):
+    """A named LLM connection configuration that can be reused across tasks."""
+
+    profile_id: str
+    name: str
+    model: str
+    base_url: str
+    api_key: str
+    timeout_seconds: int = 30
+    created_at: str = Field(default_factory=_utcnow_iso)
+    updated_at: str = Field(default_factory=_utcnow_iso)
+
+
+class CreateProfileRequest(BaseModel):
+    """Request body for creating or updating an LLM profile."""
+
+    name: str
+    model: str
+    base_url: str
+    api_key: str
+    timeout_seconds: int = 30
+
+
+class ProfileApplyRequest(BaseModel):
+    """Request body to patch LLM profile selections into a scenario YAML."""
+
+    scenario_path: str
+    judge_profile_id: str | None = None
+    answer_profile_id: str | None = None
+    dataset_profile_id: str | None = None
+
+
+class ProfileApplyResponse(BaseModel):
+    """Response after patching a scenario YAML with profile settings."""
+
+    scenario_path: str
+    patched_fields: list[str] = Field(default_factory=list)
+
+
 def jsonable(value: Any) -> Any:
     """Convert NaN/inf floats into None so the payload stays valid JSON."""
     import math

webapp/server.py

@@ -13,7 +13,7 @@ from fastapi import FastAPI
 from fastapi.responses import FileResponse
 from fastapi.staticfiles import StaticFiles
 
-from webapp.api import evaluations, runs, scenarios
+from webapp.api import evaluations, llm_profiles, runs, scenarios
 
 STATIC_DIR = Path(__file__).resolve().parent / "static"
 
@@ -29,6 +29,7 @@ def create_app() -> FastAPI:
     app.include_router(runs.router)
     app.include_router(scenarios.router)
     app.include_router(evaluations.router)
+    app.include_router(llm_profiles.router)
 
     @app.get("/api/health", tags=["meta"])
     def health() -> dict[str, str]:

webapp/services/profile_manager.py

@@ -0,0 +1,137 @@
+"""In-memory + JSON-file LLM profile manager.
+
+Profiles are kept in a dict keyed by profile_id and written to a JSON file
+on every mutation, so they survive server restarts. The pattern mirrors
+TaskManager but without threading concerns beyond a simple lock (profiles
+are only mutated by API calls in FastAPI request handlers).
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+import uuid
+from datetime import datetime, timezone
+from pathlib import Path
+
+from webapp.models import LLMProfile
+
+
+_DEFAULT_STORE = Path(__file__).resolve().parents[2] / "configs" / "llm_profiles.json"
+
+
+def _now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+class ProfileManager:
+    """Manages LLM profiles with in-memory cache and JSON file persistence."""
+
+    def __init__(self, store_path: Path = _DEFAULT_STORE) -> None:
+        self._store_path = store_path
+        self._lock = threading.Lock()
+        self._profiles: dict[str, LLMProfile] = {}
+        self._load()
+
+    # ------------------------------------------------------------------ #
+    # Public API
+    # ------------------------------------------------------------------ #
+
+    def list_all(self) -> list[LLMProfile]:
+        """Return all profiles sorted by creation time."""
+        with self._lock:
+            return sorted(self._profiles.values(), key=lambda p: p.created_at)
+
+    def get(self, profile_id: str) -> LLMProfile | None:
+        """Return one profile by id, or None if not found."""
+        with self._lock:
+            return self._profiles.get(profile_id)
+
+    def create(
+        self,
+        name: str,
+        model: str,
+        base_url: str,
+        api_key: str,
+        timeout_seconds: int = 30,
+    ) -> LLMProfile:
+        """Create and persist a new profile, returning it."""
+        now = _now_iso()
+        profile = LLMProfile(
+            profile_id=uuid.uuid4().hex[:12],
+            name=name,
+            model=model,
+            base_url=base_url,
+            api_key=api_key,
+            timeout_seconds=timeout_seconds,
+            created_at=now,
+            updated_at=now,
+        )
+        with self._lock:
+            self._profiles[profile.profile_id] = profile
+            self._persist()
+        return profile
+
+    def update(
+        self,
+        profile_id: str,
+        name: str,
+        model: str,
+        base_url: str,
+        api_key: str,
+        timeout_seconds: int = 30,
+    ) -> LLMProfile | None:
+        """Update an existing profile in-place; returns None if not found."""
+        with self._lock:
+            existing = self._profiles.get(profile_id)
+            if existing is None:
+                return None
+            updated = existing.model_copy(update={
+                "name": name,
+                "model": model,
+                "base_url": base_url,
+                "api_key": api_key,
+                "timeout_seconds": timeout_seconds,
+                "updated_at": _now_iso(),
+            })
+            self._profiles[profile_id] = updated
+            self._persist()
+        return updated
+
+    def delete(self, profile_id: str) -> bool:
+        """Remove a profile; returns True if deleted, False if not found."""
+        with self._lock:
+            if profile_id not in self._profiles:
+                return False
+            del self._profiles[profile_id]
+            self._persist()
+        return True
+
+    # ------------------------------------------------------------------ #
+    # Persistence helpers
+    # ------------------------------------------------------------------ #
+
+    def _load(self) -> None:
+        """Load profiles from the JSON store file, ignoring missing/corrupt files."""
+        if not self._store_path.exists():
+            return
+        try:
+            data = json.loads(self._store_path.read_text(encoding="utf-8"))
+            for raw in data.get("profiles", []):
+                p = LLMProfile.model_validate(raw)
+                self._profiles[p.profile_id] = p
+        except Exception:  # noqa: BLE001
+            pass  # Corrupt store — start fresh
+
+    def _persist(self) -> None:
+        """Write current profiles to the JSON store file (called under lock)."""
+        self._store_path.parent.mkdir(parents=True, exist_ok=True)
+        payload = {"profiles": [p.model_dump() for p in self._profiles.values()]}
+        self._store_path.write_text(
+            json.dumps(payload, ensure_ascii=False, indent=2),
+            encoding="utf-8",
+        )
+
+
+# Module-level singleton shared by FastAPI routes.
+profile_manager = ProfileManager()

webapp/services/report_builder.py

@@ -164,12 +164,14 @@ def build_report(run_dir: Path, metrics: list[str]) -> ReportData:
     """Build the full aggregated report payload for one run directory."""
     frame = run_reader.read_scores_frame(run_dir)
     summary_markdown = run_reader.read_summary_markdown(run_dir)
+    advice_markdown = run_reader.read_advice_markdown(run_dir)
 
     if frame.empty or not metrics:
         return ReportData(
             metrics=metrics,
             metric_means={metric: None for metric in metrics},
             summary_markdown=summary_markdown,
+            advice_markdown=advice_markdown,
         )
 
     distributions = {
@@ -185,4 +187,5 @@ def build_report(run_dir: Path, metrics: list[str]) -> ReportData:
         groupings=_groupings(frame, metrics),
         lowest_samples=_lowest_samples(frame, metrics),
         summary_markdown=summary_markdown,
+        advice_markdown=advice_markdown,
     )

webapp/services/run_reader.py

@@ -220,3 +220,14 @@ def read_summary_markdown(run_dir: Path) -> str:
         return summary_path.read_text(encoding="utf-8")
     except OSError:
         return ""
+
+
+def read_advice_markdown(run_dir: Path) -> str:
+    """Return the optimization_advice.md for a run, or an empty string if not generated."""
+    advice_path = run_dir / "optimization_advice.md"
+    if not advice_path.is_file():
+        return ""
+    try:
+        return advice_path.read_text(encoding="utf-8")
+    except OSError:
+        return ""

webapp/services/yaml_patcher.py

@@ -0,0 +1,74 @@
+"""Patch LLM profile settings into scenario YAML files in-place.
+
+Only the fields that correspond to a provided (non-None) profile are touched.
+All other fields and structure are preserved as much as PyYAML allows
+(comments are lost on round-trip, which is an accepted trade-off).
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+from webapp.models import LLMProfile
+
+
+def _repo_root() -> Path:
+    return Path(__file__).resolve().parents[2]
+
+
+def _resolve_scenario_path(path_str: str) -> Path:
+    """Resolve a scenario path; absolute paths are used as-is."""
+    candidate = Path(path_str)
+    if candidate.is_absolute():
+        return candidate
+    return (_repo_root() / candidate).resolve()
+
+
+def apply_profiles_to_scenario(
+    scenario_path: str,
+    judge_profile: LLMProfile | None,
+    answer_profile: LLMProfile | None,
+    dataset_profile: LLMProfile | None,
+    _resolve_absolute: bool = False,
+) -> list[str]:
+    """Patch the YAML file at *scenario_path* with the supplied profiles.
+
+    Returns a list of dotted field names that were actually patched.
+    Setting *_resolve_absolute=True* skips repo-root resolution (used in tests).
+    """
+    if _resolve_absolute:
+        resolved = Path(scenario_path)
+    else:
+        resolved = _resolve_scenario_path(scenario_path)
+
+    if not resolved.exists():
+        raise FileNotFoundError(f"Scenario file not found: {resolved}")
+
+    data: dict[str, Any] = yaml.safe_load(resolved.read_text(encoding="utf-8")) or {}
+    patched: list[str] = []
+
+    if judge_profile is not None:
+        data["judge_model"] = judge_profile.model
+        patched.append("judge_model")
+
+    if answer_profile is not None:
+        adapter = data.get("app_adapter")
+        if isinstance(adapter, dict):
+            static_kwargs = adapter.setdefault("static_kwargs", {})
+            static_kwargs["model"] = answer_profile.model
+            patched.append("app_adapter.static_kwargs.model")
+
+    if dataset_profile is not None:
+        generation = data.get("generation")
+        if isinstance(generation, dict):
+            generation["model"] = dataset_profile.model
+            patched.append("generation.model")
+
+    resolved.write_text(
+        yaml.dump(data, allow_unicode=True, default_flow_style=False, sort_keys=False),
+        encoding="utf-8",
+    )
+    return patched

webapp/static/css/app.css

@@ -265,3 +265,69 @@ table.group-table td { border-bottom: 1px solid #f1f5f9; font-variant-numeric: t
   .sidebar { width: 64px; }
   .brand-sub, .nav-item span:not(.nav-ico), .sidebar-foot span:last-child { display: none; }
 }
+
+/* ---------- LLM 配置管理页 ---------- */
+.profile-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(300px, 1fr)); gap: 16px; }
+.profile-card {
+  background: var(--surface); border: 1px solid var(--line); border-radius: var(--radius);
+  padding: 16px; box-shadow: var(--shadow);
+}
+.profile-card-head { display: flex; justify-content: space-between; align-items: center; margin-bottom: 10px; }
+.profile-card-name { font-size: 15px; font-weight: 600; }
+.profile-card-actions { display: flex; gap: 6px; }
+.profile-card-field { font-size: 12px; color: var(--slate); margin-top: 4px; }
+.field-label { font-weight: 600; color: var(--ink); }
+
+/* Form */
+.profile-form { display: flex; flex-direction: column; gap: 12px; margin-top: 14px; max-width: 560px; }
+.form-row { display: flex; flex-direction: column; gap: 4px; }
+.form-label { font-size: 13px; font-weight: 600; }
+.req { color: var(--bad); }
+.form-input {
+  border: 1px solid var(--line); border-radius: 6px; padding: 8px 10px;
+  font-size: 13px; font-family: inherit; width: 100%;
+}
+.form-input:focus { outline: none; border-color: var(--petrol); }
+.form-input-sm { max-width: 120px; }
+.form-actions { display: flex; gap: 10px; align-items: center; margin-top: 4px; }
+.form-error { font-size: 12px; color: var(--bad); }
+.btn-sm { padding: 4px 10px; font-size: 12px; }
+.btn-danger { color: var(--bad); border-color: var(--bad); }
+.btn-danger:hover { background: #fee2e2; }
+
+/* 选中态 run 卡片 */
+.run-card.selected {
+  border-color: var(--petrol);
+  box-shadow: 0 0 0 2px rgba(0,153,153,0.25), var(--shadow);
+}
+
+/* ---------- LLM 角色配置面板 ---------- */
+.llm-assignment-panel { border-left: 3px solid var(--petrol); }
+.llm-role-rows { display: flex; flex-direction: column; gap: 10px; }
+.llm-role-row { display: flex; align-items: center; gap: 14px; }
+.llm-role-label { font-size: 13px; font-weight: 600; min-width: 180px; color: var(--ink); }
+.llm-role-select { min-width: 240px; }
+
+/* ---------- ⑤ 优化建议面板 ---------- */
+.advice-panel { border-left: 3px solid #7c3aed; }
+.advice-header {
+  display: flex; align-items: center; gap: 10px;
+  margin-bottom: 14px;
+}
+.advice-badge {
+  background: #7c3aed; color: #fff;
+  font-size: 11px; font-weight: 700; letter-spacing: 0.5px;
+  padding: 3px 8px; border-radius: 4px; text-transform: uppercase;
+}
+.advice-model { font-size: 12px; color: var(--slate); }
+.advice-body { line-height: 1.7; color: var(--ink); }
+.advice-md h1 { font-size: 16px; font-weight: 700; margin: 16px 0 8px; color: var(--ink); }
+.advice-md h2 {
+  font-size: 14px; font-weight: 700; margin: 20px 0 8px;
+  padding-bottom: 4px; border-bottom: 1px solid var(--line); color: var(--ink-soft);
+}
+.advice-md h3 { font-size: 13px; font-weight: 600; margin: 12px 0 6px; color: var(--slate); }
+.advice-md hr { border: none; border-top: 1px solid var(--line); margin: 14px 0; }
+.advice-md ul { padding-left: 20px; margin: 6px 0; }
+.advice-md li { margin: 3px 0; font-size: 13px; }
+.advice-md strong { color: var(--ink); font-weight: 600; }

webapp/static/index.html

@@ -22,9 +22,12 @@
         <button class="nav-item" data-view="new">
           <span class="nav-ico">＋</span><span>新建评估</span>
         </button>
-        <button class="nav-item" data-view="report" data-requires-run="1">
+        <button class="nav-item" data-view="report" data-requires-run="1" disabled>
           <span class="nav-ico">▤</span><span>报告详情</span>
         </button>
+        <button class="nav-item" data-view="profiles">
+          <span class="nav-ico">⚙</span><span>LLM 配置</span>
+        </button>
       </nav>
       <div class="sidebar-foot">
         <span class="dot" id="health-dot"></span>
@@ -59,6 +62,33 @@
             <span class="selected-scenario muted" id="selected-scenario">未选择场景</span>
           </div>
         </div>
+
+        <!-- LLM 角色配置面板（选中场景后显示） -->
+        <div class="panel llm-assignment-panel" id="llm-assignment-panel" hidden>
+          <h2>LLM 角色配置 <span class="muted" style="font-size:13px;font-weight:400">（可选）</span></h2>
+          <p class="muted" style="margin-bottom:14px">为不同任务角色选择已保存的 LLM 配置，留空则使用场景文件中的原始配置。</p>
+          <div class="llm-role-rows">
+            <div class="llm-role-row">
+              <label class="llm-role-label">评测打分 Judge LLM</label>
+              <select class="select llm-role-select" id="role-judge">
+                <option value="">— 使用场景原始配置 —</option>
+              </select>
+            </div>
+            <div class="llm-role-row">
+              <label class="llm-role-label">生成答案 Answer LLM</label>
+              <select class="select llm-role-select" id="role-answer">
+                <option value="">— 使用场景原始配置 —</option>
+              </select>
+            </div>
+            <div class="llm-role-row">
+              <label class="llm-role-label">生成题库 Dataset LLM</label>
+              <select class="select llm-role-select" id="role-dataset">
+                <option value="">— 使用场景原始配置 —</option>
+              </select>
+            </div>
+          </div>
+        </div>
+
         <div class="panel" id="task-panel" hidden>
           <div class="task-head">
             <h2>评估进度</h2>
@@ -105,6 +135,68 @@
           <!-- ④ 最低分样本逐条复核 -->
           <div class="section-label">④ 最低分样本（点击展开逐条复核）</div>
           <div class="lowest-table" id="lowest-table"></div>
+
+          <!-- ⑤ 优化建议（optimization_advisor: true 时显示） -->
+          <div id="advice-section" hidden>
+            <div class="section-label">⑤ 优化建议 OPTIMIZATION ADVICE</div>
+            <div class="panel advice-panel">
+              <div class="advice-header">
+                <span class="advice-badge">AI 诊断报告</span>
+                <span class="advice-model" id="advice-model-label"></span>
+              </div>
+              <div class="advice-body" id="advice-body"></div>
+            </div>
+          </div>
+        </div>
+      </section>
+
+      <!-- LLM 配置视图 -->
+      <section class="view" id="view-profiles" hidden>
+        <div class="panel">
+          <div class="panel-head">
+            <h2>LLM 配置管理</h2>
+            <button class="btn btn-primary" id="add-profile-btn">＋ 新建配置</button>
+          </div>
+          <p class="muted">保存常用 LLM 连接参数，在运行评估时按角色选择。</p>
+        </div>
+
+        <!-- 新建 / 编辑表单（默认隐藏） -->
+        <div class="panel" id="profile-form-panel" hidden>
+          <h2 id="profile-form-title">新建 LLM 配置</h2>
+          <div class="profile-form">
+            <input type="hidden" id="edit-profile-id" />
+            <div class="form-row">
+              <label class="form-label">配置名称 <span class="req">*</span></label>
+              <input class="form-input" id="pf-name" placeholder="例：DeepSeek Flash（内网）" />
+            </div>
+            <div class="form-row">
+              <label class="form-label">模型名称 <span class="req">*</span></label>
+              <input class="form-input" id="pf-model" placeholder="例：deepseek-v4-flash" />
+            </div>
+            <div class="form-row">
+              <label class="form-label">Base URL <span class="req">*</span></label>
+              <input class="form-input" id="pf-base-url" placeholder="例：http://6.86.80.4:30080/v1" />
+            </div>
+            <div class="form-row">
+              <label class="form-label">API Key <span class="req">*</span></label>
+              <input class="form-input" id="pf-api-key" type="password" placeholder="sk-…" />
+            </div>
+            <div class="form-row">
+              <label class="form-label">超时（秒）</label>
+              <input class="form-input form-input-sm" id="pf-timeout" type="number" value="30" min="5" max="300" />
+            </div>
+            <div class="form-actions">
+              <button class="btn btn-primary" id="save-profile-btn">保存</button>
+              <button class="btn" id="cancel-profile-btn">取消</button>
+              <span class="form-error muted" id="profile-form-error"></span>
+            </div>
+          </div>
+        </div>
+
+        <div id="profile-cards" class="profile-grid"></div>
+        <div class="empty" id="profiles-empty" hidden>
+          <p>尚未添加任何 LLM 配置。</p>
+          <p class="muted">点击「新建配置」添加第一个。</p>
         </div>
       </section>
     </main>
@@ -112,6 +204,7 @@
 
   <script src="/static/js/api.js"></script>
   <script src="/static/js/report.js"></script>
+  <script src="/static/js/profiles.js"></script>
   <script src="/static/js/runner.js"></script>
   <script src="/static/js/app.js"></script>
 </body>

webapp/static/js/api.js

@@ -43,4 +43,26 @@ const API = {
     return API.post("/api/evaluations", { scenario_path: scenarioPath });
   },
   taskStatus(taskId) { return API.get(`/api/evaluations/${encodeURIComponent(taskId)}`); },
+
+  // LLM Profile API
+  profiles() { return API.get("/api/llm-profiles"); },
+  createProfile(body) { return API.post("/api/llm-profiles", body); },
+  updateProfile(id, body) {
+    return fetch(`/api/llm-profiles/${encodeURIComponent(id)}`, {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    }).then(async r => {
+      if (!r.ok) { const d = await API._extractError(r); throw new Error(d); }
+      return r.json();
+    });
+  },
+  deleteProfile(id) {
+    return fetch(`/api/llm-profiles/${encodeURIComponent(id)}`, { method: "DELETE" })
+      .then(async r => {
+        if (!r.ok) { const d = await API._extractError(r); throw new Error(d); }
+        return r.json();
+      });
+  },
+  applyProfiles(body) { return API.post("/api/llm-profiles/apply", body); },
 };

webapp/static/js/app.js

@@ -1,28 +1,59 @@
 // app.js — 视图路由、运行列表渲染、健康检查。整个控制台的入口编排。
+// 会话保持：URL hash 路由（#runs / #new / #profiles / #report/{runId}）
+//            + sessionStorage 兜底，F5 刷新 / 浏览器前进后退均可恢复。
 
 const App = {
   currentRunId: null,
-  views: ["runs", "new", "report"],
+  activeView: null,
-  titles: { runs: "运行列表", new: "新建评估", report: "报告详情" },
+  views: ["runs", "new", "report", "profiles"],
+  titles: { runs: "运行列表", new: "新建评估", report: "报告详情", profiles: "LLM 配置" },
 
-  // 初始化：绑定导航、加载首屏、启动健康检查。
+  // 初始化：绑定导航、从 URL/sessionStorage 恢复上次位置、启动健康检查。
   init() {
     document.querySelectorAll(".nav-item").forEach((btn) => {
-      btn.addEventListener("click", () => App.switchView(btn.dataset.view));
+      btn.addEventListener("click", () => App.navigate(btn.dataset.view));
     });
     document.getElementById("refresh-btn").addEventListener("click", () => App.refreshCurrent());
 
     Runner.init();
-    App.switchView("runs");
+    Profiles.init();
+
+    // 恢复上次会话（优先 URL hash，其次 sessionStorage）
+    App._restoreSession();
+
     App.checkHealth();
     setInterval(App.checkHealth, 15000);
+
+    // 浏览器前进 / 后退按钮
+    window.addEventListener("popstate", () => App._restoreSession());
   },
 
-  // 切换主视图，并同步导航高亮与标题。
+  // ----------------------------------------------------------------
-  switchView(view) {
+  // 路由 —— 有历史记录的主动导航（更新 URL hash）
-    if (view === "report" && !App.currentRunId) {
+  // ----------------------------------------------------------------
-      // 没有选中的运行时，报告页显示占位。
+  navigate(view, runId) {
+    if (runId !== undefined) App.currentRunId = runId;
+    const hash = App._buildHash(view, App.currentRunId);
+    if (location.hash !== `#${hash}`) {
+      history.pushState({ view, runId: App.currentRunId }, "", `#${hash}`);
     }
+    App._doSwitch(view);
+  },
+
+  // 供内部调用（不产生历史记录），例如刷新同一视图
+  switchView(view) {
+    App._doSwitch(view);
+  },
+
+  // 刷新当前视图数据
+  refreshCurrent() {
+    App._doSwitch(App.activeView || "runs");
+  },
+
+  // ----------------------------------------------------------------
+  // 内部：实际切换 DOM + 触发数据加载
+  // ----------------------------------------------------------------
+  _doSwitch(view) {
     App.views.forEach((name) => {
       const el = document.getElementById(`view-${name}`);
       if (el) el.hidden = name !== view;
@@ -33,17 +64,53 @@ const App = {
     document.getElementById("view-title").textContent = App.titles[view] || view;
     App.activeView = view;
 
+    // 持久化到 sessionStorage（URL 共享场景的备份）
+    sessionStorage.setItem("rag_view", view);
+    if (App.currentRunId) sessionStorage.setItem("rag_run_id", App.currentRunId);
+
     if (view === "runs")     App.loadRuns();
     if (view === "new")      Runner.loadScenarios();
     if (view === "report")   Report.render(App.currentRunId);
+    if (view === "profiles") Profiles.load();
   },
 
-  // 刷新当前视图的数据。
+  // ----------------------------------------------------------------
-  refreshCurrent() {
+  // Hash 工具
-    App.switchView(App.activeView || "runs");
+  // ----------------------------------------------------------------
+  _buildHash(view, runId) {
+    if (view === "report" && runId) {
+      return `report/${encodeURIComponent(runId)}`;
+    }
+    return view || "runs";
   },
 
-  // 加载并渲染运行列表。
+  _parseHash() {
+    const raw = location.hash.replace(/^#\/?/, "");
+    if (!raw) return { view: null, runId: null };
+    if (raw.startsWith("report/")) {
+      const runId = decodeURIComponent(raw.slice("report/".length));
+      return { view: "report", runId };
+    }
+    const view = App.views.includes(raw) ? raw : null;
+    return { view, runId: null };
+  },
+
+  // 会话恢复：hash → sessionStorage → 默认 runs
+  _restoreSession() {
+    const { view: hView, runId: hRunId } = App._parseHash();
+    const view   = hView  || sessionStorage.getItem("rag_view")   || "runs";
+    const runId  = hRunId || sessionStorage.getItem("rag_run_id") || null;
+
+    if (runId) {
+      App.currentRunId = runId;
+      App.enableReportNav();
+    }
+    App._doSwitch(view);
+  },
+
+  // ----------------------------------------------------------------
+  // 运行列表
+  // ----------------------------------------------------------------
   async loadRuns() {
     const container = document.getElementById("runs-container");
     const empty = document.getElementById("runs-empty");
@@ -64,14 +131,16 @@ const App = {
     }
   },
 
-  // 构造一张运行卡片。
   renderRunCard(run) {
     const card = document.createElement("div");
-    card.className = "run-card";
+    card.className = "run-card" + (run.run_id === App.currentRunId ? " selected" : "");
+
     card.addEventListener("click", () => {
-      App.currentRunId = run.run_id;
+      // 更新选中高亮
+      document.querySelectorAll(".run-card").forEach((c) => c.classList.remove("selected"));
+      card.classList.add("selected");
       App.enableReportNav();
-      App.switchView("report");
+      App.navigate("report", run.run_id);
     });
 
     const chips = (run.metrics || [])
@@ -79,7 +148,7 @@ const App = {
         const val = run.metric_means ? run.metric_means[m] : null;
         const cls = App.scoreClass(val);
         const text = val === null || val === undefined ? "n/a" : val.toFixed(2);
-        return `<span class="metric-chip">${App.escape(App.shortMetric(m))} <b class="${cls}">${text}</b></span>`;
+        return `<span class="metric-chip" title="${App.escape(m)}">${App.escape(App.shortMetric(m))} <b class="${cls}">${text}</b></span>`;
       })
       .join("");
 
@@ -96,13 +165,14 @@ const App = {
     return card;
   },
 
-  // 启用报告导航项（选中运行后）。
+  // ----------------------------------------------------------------
+  // 工具方法
+  // ----------------------------------------------------------------
   enableReportNav() {
     const btn = document.querySelector('.nav-item[data-view="report"]');
     if (btn) btn.disabled = false;
   },
 
-  // 根据分值返回 good/warn/bad/na 配色类。
   scoreClass(value) {
     if (value === null || value === undefined) return "na";
     if (value >= 0.8) return "good";
@@ -110,31 +180,30 @@ const App = {
     return "bad";
   },
 
-  // 指标名缩写，节省卡片横向空间。
   shortMetric(name) {
     const map = {
       faithfulness:        "faith.",
       answer_relevancy:    "ans.rel.",
       context_recall:      "ctx.recall",
       context_precision:   "ctx.prec.",
+      noise_sensitivity:   "noise.sens.",
+      factual_correctness: "fact.corr.",
+      semantic_similarity: "sem.sim.",
     };
     return map[name] || name;
   },
 
-  // 截取时间戳到分钟，便于阅读。
   shortTime(iso) {
     if (!iso) return "—";
     return String(iso).replace("T", " ").slice(0, 16);
   },
 
-  // 简单 HTML 转义，防止注入。
   escape(text) {
     const div = document.createElement("div");
     div.textContent = text == null ? "" : String(text);
     return div.innerHTML;
   },
 
-  // 健康检查，更新左下角状态点。
   async checkHealth() {
     const dot   = document.getElementById("health-dot");
     const label = document.getElementById("health-text");

webapp/static/js/profiles.js

@@ -0,0 +1,118 @@
+// profiles.js — LLM 配置管理页面逻辑
+
+const Profiles = {
+  _data: [],
+
+  // 初始化：绑定按钮事件
+  init() {
+    document.getElementById("add-profile-btn").addEventListener("click", () => Profiles.showForm());
+    document.getElementById("save-profile-btn").addEventListener("click", () => Profiles.save());
+    document.getElementById("cancel-profile-btn").addEventListener("click", () => Profiles.hideForm());
+  },
+
+  // 加载并渲染 Profile 列表
+  async load() {
+    const grid = document.getElementById("profile-cards");
+    const empty = document.getElementById("profiles-empty");
+    grid.innerHTML = '<p class="muted">加载中…</p>';
+    try {
+      const data = await API.profiles();
+      Profiles._data = data.profiles || [];
+      grid.innerHTML = "";
+      if (Profiles._data.length === 0) {
+        empty.hidden = false;
+      } else {
+        empty.hidden = true;
+        Profiles._data.forEach(p => grid.appendChild(Profiles.renderCard(p)));
+      }
+    } catch (err) {
+      grid.innerHTML = `<p class="muted">加载失败：${App.escape(err.message)}</p>`;
+    }
+  },
+
+  // 渲染单个 Profile 卡片
+  renderCard(p) {
+    const card = document.createElement("div");
+    card.className = "profile-card";
+    card.dataset.id = p.profile_id;
+    card.innerHTML = `
+      <div class="profile-card-head">
+        <div class="profile-card-name">${App.escape(p.name)}</div>
+        <div class="profile-card-actions">
+          <button class="btn btn-sm" data-action="edit">编辑</button>
+          <button class="btn btn-sm btn-danger" data-action="delete">删除</button>
+        </div>
+      </div>
+      <div class="profile-card-field"><span class="field-label">模型</span> <code>${App.escape(p.model)}</code></div>
+      <div class="profile-card-field"><span class="field-label">Base URL</span> <code>${App.escape(p.base_url)}</code></div>
+      <div class="profile-card-field"><span class="field-label">超时</span> ${p.timeout_seconds}s</div>
+    `;
+    card.querySelector("[data-action=edit]").addEventListener("click", () => Profiles.showForm(p));
+    card.querySelector("[data-action=delete]").addEventListener("click", () => Profiles.remove(p.profile_id, p.name));
+    return card;
+  },
+
+  // 显示新建或编辑表单
+  showForm(profile = null) {
+    const panel = document.getElementById("profile-form-panel");
+    const title = document.getElementById("profile-form-title");
+    panel.hidden = false;
+    title.textContent = profile ? "编辑 LLM 配置" : "新建 LLM 配置";
+    document.getElementById("edit-profile-id").value = profile ? profile.profile_id : "";
+    document.getElementById("pf-name").value = profile ? profile.name : "";
+    document.getElementById("pf-model").value = profile ? profile.model : "";
+    document.getElementById("pf-base-url").value = profile ? profile.base_url : "";
+    document.getElementById("pf-api-key").value = profile ? profile.api_key : "";
+    document.getElementById("pf-timeout").value = profile ? profile.timeout_seconds : 30;
+    document.getElementById("profile-form-error").textContent = "";
+    panel.scrollIntoView({ behavior: "smooth", block: "start" });
+  },
+
+  hideForm() {
+    document.getElementById("profile-form-panel").hidden = true;
+  },
+
+  // 保存（新建 or 更新）
+  async save() {
+    const id = document.getElementById("edit-profile-id").value;
+    const body = {
+      name: document.getElementById("pf-name").value.trim(),
+      model: document.getElementById("pf-model").value.trim(),
+      base_url: document.getElementById("pf-base-url").value.trim(),
+      api_key: document.getElementById("pf-api-key").value.trim(),
+      timeout_seconds: parseInt(document.getElementById("pf-timeout").value, 10) || 30,
+    };
+    const errEl = document.getElementById("profile-form-error");
+    if (!body.name || !body.model || !body.base_url || !body.api_key) {
+      errEl.textContent = "请填写所有必填字段（名称、模型、Base URL、API Key）";
+      return;
+    }
+    try {
+      if (id) {
+        await API.updateProfile(id, body);
+      } else {
+        await API.createProfile(body);
+      }
+      Profiles.hideForm();
+      await Profiles.load();
+    } catch (err) {
+      errEl.textContent = `保存失败：${err.message}`;
+    }
+  },
+
+  // 删除 Profile
+  async remove(profileId, name) {
+    if (!confirm(`确认删除配置「${name}」？`)) return;
+    try {
+      await API.deleteProfile(profileId);
+      await Profiles.load();
+    } catch (err) {
+      alert(`删除失败：${err.message}`);
+    }
+  },
+
+  // 获取当前已加载的 profiles（供 runner.js 使用）
+  getAll() {
+    return Profiles._data;
+  },
+};

webapp/static/js/report.js

@@ -26,6 +26,7 @@ const Report = {
       Report.renderDistribution(detail.report);
       Report.renderGroupings(detail.report);
       Report.renderLowest(detail.report);
+      Report.renderAdvice(detail.summary, detail.report);
       content.style.opacity = "1";
     } catch (err) {
       empty.hidden = false;
@@ -186,8 +187,7 @@ const Report = {
   },
 
   // ④ 最低分样本逐条复核表（点击展开）。
-  renderLowest(report) {
+  renderLowest(report) {    const wrap = document.getElementById("lowest-table");
-    const wrap = document.getElementById("lowest-table");
     const samples = report.lowest_samples || [];
     wrap.innerHTML = "";
     if (samples.length === 0) {
@@ -255,4 +255,35 @@ const Report = {
       </div>
     `;
   },
+
+  // ⑤ 优化建议（仅 optimization_advice.md 存在时渲染）。
+  renderAdvice(summary, report) {
+    const section = document.getElementById("advice-section");
+    const body = document.getElementById("advice-body");
+    const modelLabel = document.getElementById("advice-model-label");
+
+    const md = report.advice_markdown || "";
+    if (!md.trim()) {
+      section.hidden = true;
+      return;
+    }
+
+    section.hidden = false;
+    modelLabel.textContent = summary.judge_model ? `judge: ${summary.judge_model}` : "";
+
+    // 简单 Markdown → HTML 转换（标题、列表、分隔线、粗体）
+    const escaped = md
+      .replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
+    const html = escaped
+      .replace(/^#{3}\s+(.+)$/gm, "<h3>$1</h3>")
+      .replace(/^#{2}\s+(.+)$/gm, "<h2>$1</h2>")
+      .replace(/^#{1}\s+(.+)$/gm, "<h1>$1</h1>")
+      .replace(/^---+$/gm, "<hr>")
+      .replace(/\*\*(.+?)\*\*/g, "<strong>$1</strong>")
+      .replace(/^- (.+)$/gm, "<li>$1</li>")
+      .replace(/(<li>[^]*?<\/li>\n?)+/g, (m) => `<ul>${m}</ul>`)
+      .replace(/\n\n+/g, "\n<br>\n");
+
+    body.innerHTML = `<div class="advice-md">${html}</div>`;
+  },
 };

webapp/static/js/runner.js

@@ -1,17 +1,17 @@
-// runner.js — 新建评估视图：列出场景、触发评估、轮询任务状态与日志。
+// runner.js — 新建评估视图：列出场景、LLM角色配置、触发评估、轮询任务状态与日志。
 
 const Runner = {
   selectedScenario: null,
   pollTimer: null,
+  lastRunId: null,
 
   // 绑定运行按钮。
   init() {
     document.getElementById("run-btn").addEventListener("click", () => Runner.trigger());
     document.getElementById("view-report-btn").addEventListener("click", () => {
       if (Runner.lastRunId) {
-        App.currentRunId = Runner.lastRunId;
         App.enableReportNav();
-        App.switchView("report");
+        App.navigate("report", Runner.lastRunId);
       }
     });
   },
@@ -32,6 +32,27 @@ const Runner = {
     } catch (err) {
       list.innerHTML = `<p class="muted">加载失败：${App.escape(err.message)}</p>`;
     }
+    // 同时加载 profiles 供角色选择
+    Runner._populateProfileSelects();
+  },
+
+  // 填充三个角色下拉框
+  async _populateProfileSelects() {
+    const cached = Profiles.getAll();
+    const profiles = cached.length > 0
+      ? cached
+      : (await API.profiles().catch(() => ({ profiles: [] }))).profiles;
+
+    ["role-judge", "role-answer", "role-dataset"].forEach(id => {
+      const sel = document.getElementById(id);
+      sel.innerHTML = '<option value="">— 使用场景原始配置 —</option>';
+      profiles.forEach(p => {
+        const opt = document.createElement("option");
+        opt.value = p.profile_id;
+        opt.textContent = `${p.name}  (${p.model})`;
+        sel.appendChild(opt);
+      });
+    });
   },
 
   // 构造单个场景条目。
@@ -64,12 +85,14 @@ const Runner = {
         Runner.selectedScenario = sc.path;
         document.getElementById("selected-scenario").textContent = sc.path;
         document.getElementById("run-btn").disabled = false;
+        // 显示 LLM 角色面板
+        document.getElementById("llm-assignment-panel").hidden = false;
       });
     }
     return item;
   },
 
-  // 触发评估并开始轮询。
+  // 触发评估：先 apply profiles（若选了），再触发任务。
   async trigger() {
     if (!Runner.selectedScenario) return;
     const runBtn = document.getElementById("run-btn");
@@ -85,15 +108,41 @@ const Runner = {
     Runner._setStatus(statusBadge, "queued");
 
     try {
+      // Step 1: apply LLM profiles to YAML if any selected
+      await Runner._applyProfilesIfNeeded(logBox);
+
+      // Step 2: trigger evaluation
       const resp = await API.triggerEvaluation(Runner.selectedScenario);
       Runner.poll(resp.task_id);
     } catch (err) {
       Runner._setStatus(statusBadge, "failed");
-      logBox.textContent = `触发失败：${err.message}`;
+      logBox.textContent = (logBox.textContent ? logBox.textContent + "\n" : "") + `触发失败：${err.message}`;
       runBtn.disabled = false;
     }
   },
 
+  // 如果用户选了 profile，就先 apply 写回 YAML
+  async _applyProfilesIfNeeded(logBox) {
+    const judgeId = document.getElementById("role-judge").value;
+    const answerId = document.getElementById("role-answer").value;
+    const datasetId = document.getElementById("role-dataset").value;
+
+    if (!judgeId && !answerId && !datasetId) return; // 全空，跳过
+
+    logBox.textContent = "正在将 LLM 配置写入场景文件…\n";
+    const body = {
+      scenario_path: Runner.selectedScenario,
+      judge_profile_id: judgeId || null,
+      answer_profile_id: answerId || null,
+      dataset_profile_id: datasetId || null,
+    };
+    const result = await API.applyProfiles(body);
+    const fields = (result.patched_fields || []).join(", ");
+    logBox.textContent += fields
+      ? `✓ 已更新字段：${fields}\n`
+      : "（未找到可更新的字段，继续运行）\n";
+  },
+
   // 周期性轮询任务状态，刷新日志与徽标。
   poll(taskId) {
     const logBox = document.getElementById("task-log");
@@ -114,6 +163,7 @@ const Runner = {
           runBtn.disabled = false;
           if (status.status === "completed" && status.run_id) {
             Runner.lastRunId = status.run_id;
+            sessionStorage.setItem("rag_run_id", status.run_id);
             reportBtn.hidden = false;
           }
         }