feat: implement true OPRO with Gemini-style UI

- Add true OPRO system instruction optimization (vs query rewriting)
- Implement iterative optimization with performance trajectory
- Add new OPRO API endpoints (/opro/create, /opro/generate_and_evaluate, /opro/execute)
- Create modern Gemini-style chat UI (frontend/opro.html)
- Optimize performance: reduce candidates from 20 to 10 (2x faster)
- Add model selector in UI toolbar
- Add collapsible sidebar with session management
- Add copy button for instructions
- Ensure all generated prompts use simplified Chinese
- Update README with comprehensive documentation
- Add .gitignore for local_docs folder

.gitignore

@@ -147,6 +147,7 @@ cython_debug/
 outputs/
 *.jsonl
 *.log
+local_docs/
 
 # Node modules (if any frontend dependencies)
 node_modules/

README.md

@@ -1,3 +1,218 @@
+# OPRO Prompt Optimizer
+
+## 功能概述
+
+OPRO (Optimization by PROmpting) 是一个基于大语言模型的提示词优化系统。本项目实现了真正的 OPRO 算法，通过迭代优化系统指令（System Instructions）来提升 LLM 在特定任务上的性能。
+
+### 核心功能
+
+- **系统指令优化**：使用 LLM 作为优化器，基于历史性能轨迹生成更优的系统指令
+- **多轮迭代优化**：支持多轮优化，每轮基于前一轮的性能反馈生成新的候选指令
+- **智能候选选择**：通过语义聚类和多样性选择，从大量候选中筛选出最具代表性的指令
+- **性能评估**：支持自定义测试用例对系统指令进行自动评估
+- **会话管理**：支持多个优化任务的并行管理和历史记录
+
+### 用户界面
+
+- **现代化聊天界面**：类似 Google Gemini 的简洁设计
+- **侧边栏会话管理**：可折叠的侧边栏，支持多会话切换
+- **实时优化反馈**：每轮优化生成 3-5 个候选指令，用户可选择继续优化或执行
+- **模型选择**：支持在界面中选择不同的 LLM 模型
+
+## 主要优化改进
+
+### 1. 真正的 OPRO 实现
+
+原始代码实现的是查询重写（Query Rewriting），而非真正的 OPRO。我们添加了完整的 OPRO 功能：
+
+- **系统指令生成**：`generate_system_instruction_candidates()` - 生成多样化的系统指令候选
+- **性能评估**：`evaluate_system_instruction()` - 基于测试用例评估指令性能
+- **轨迹优化**：基于历史 (instruction, score) 轨迹生成更优指令
+- **元提示工程**：专门设计的元提示用于指导 LLM 生成和优化系统指令
+
+### 2. 性能优化
+
+- **候选池大小优化**：从 20 个候选减少到 10 个，速度提升约 2 倍
+- **智能聚类选择**：使用 AgglomerativeClustering 从候选池中选择最具多样性的 Top-K
+- **嵌入服务回退**：Xinference → Ollama 自动回退机制，确保服务可用性
+
+### 3. API 架构改进
+
+- **新增 OPRO 端点**：
+  - `POST /opro/create` - 创建 OPRO 优化任务
+  - `POST /opro/generate_and_evaluate` - 生成并自动评估候选
+  - `POST /opro/execute` - 执行系统指令
+  - `GET /opro/runs` - 获取所有优化任务
+  - `GET /opro/run/{run_id}` - 获取特定任务详情
+- **会话状态管理**：完整的 OPRO 运行状态跟踪（轨迹、测试用例、迭代次数）
+- **向后兼容**：保留原有查询重写功能，标记为 `opro-legacy`
+
+### 4. 前端界面重构
+
+- **Gemini 风格设计**：简洁的白色/灰色配色，圆角设计，微妙的阴影效果
+- **可折叠侧边栏**：默认折叠，支持会话列表管理
+- **多行输入框**：支持多行文本输入，底部工具栏包含模型选择器
+- **候选指令卡片**：每个候选显示编号、内容、分数，提供"继续优化"、"复制"、"执行"按钮
+- **简体中文界面**：所有 UI 文本和生成的指令均使用简体中文
+
+## 快速开始
+
+### 环境要求
+
+- **Python** ≥ 3.10（推荐使用 conda 虚拟环境）
+- **Ollama** 本地服务及模型（如 `qwen3:8b`、`qwen3-embedding:4b`）
+- **可选**：Xinference embedding 服务
+
+### 安装依赖
+
+```bash
+# 创建 conda 环境（推荐）
+conda create -n opro python=3.10
+conda activate opro
+
+# 安装 Python 依赖
+pip install fastapi uvicorn requests numpy scikit-learn pydantic
+```
+
+### 启动 Ollama 服务
+
+```bash
+# 确保 Ollama 已安装并运行
+ollama serve
+
+# 拉取所需模型
+ollama pull qwen3:8b
+ollama pull qwen3-embedding:4b
+```
+
+### 启动应用
+
+```bash
+# 启动后端服务
+uvicorn _qwen_xinference_demo.api:app --host 127.0.0.1 --port 8010
+
+# 或使用 0.0.0.0 允许外部访问
+uvicorn _qwen_xinference_demo.api:app --host 0.0.0.0 --port 8010
+```
+
+### 访问界面
+
+- **OPRO 优化界面**：http://127.0.0.1:8010/ui/opro.html
+- **传统三栏界面**：http://127.0.0.1:8010/ui/
+- **API 文档**：http://127.0.0.1:8010/docs
+- **OpenAPI JSON**：http://127.0.0.1:8010/openapi.json
+
+### 使用示例
+
+1. **创建新会话**：在 OPRO 界面点击"新建会话"或侧边栏的 + 按钮
+2. **输入任务描述**：例如"将中文翻译成英文"
+3. **查看候选指令**：系统生成 3-5 个优化的系统指令
+4. **继续优化**：点击"继续优化"进行下一轮迭代
+5. **执行指令**：点击"执行此指令"测试指令效果
+6. **复制指令**：点击"复制"按钮将指令复制到剪贴板
+
+## 配置说明
+
+配置文件：`config.py`
+
+### 关键配置项
+
+```python
+# Ollama 服务配置
+OLLAMA_HOST = "http://127.0.0.1:11434"
+DEFAULT_CHAT_MODEL = "qwen3:8b"
+DEFAULT_EMBED_MODEL = "qwen3-embedding:4b"
+
+# OPRO 优化参数
+GENERATION_POOL_SIZE = 10  # 生成候选池大小
+TOP_K = 5                   # 返回给用户的候选数量
+CLUSTER_DISTANCE_THRESHOLD = 0.15  # 聚类距离阈值
+
+# Xinference 配置（可选）
+XINFERENCE_EMBED_URL = "http://127.0.0.1:9997/models/bge-base-zh/embed"
+```
+
+## 项目结构
+
+```
+.
+├── _qwen_xinference_demo/
+│   ├── api.py                          # FastAPI 主应用
+│   └── opro/
+│       ├── user_prompt_optimizer.py    # OPRO 核心逻辑
+│       ├── prompt_utils.py             # 元提示生成
+│       ├── session_state.py            # 会话状态管理
+│       ├── ollama_client.py            # Ollama 客户端
+│       └── xinference_client.py        # Xinference 客户端
+├── frontend/
+│   ├── opro.html                       # OPRO 优化界面
+│   └── index.html                      # 传统三栏界面
+├── examples/
+│   ├── opro_demo.py                    # OPRO 功能演示
+│   └── client_demo.py                  # API 调用示例
+├── config.py                           # 全局配置
+├── API.md                              # API 文档
+└── README.md                           # 本文件
+```
+
+## API 端点
+
+### OPRO 相关（推荐使用）
+
+- `POST /opro/create` - 创建优化任务
+- `POST /opro/generate_and_evaluate` - 生成并评估候选
+- `POST /opro/execute` - 执行系统指令
+- `GET /opro/runs` - 获取所有任务
+- `GET /opro/run/{run_id}` - 获取任务详情
+
+### 传统端点（向后兼容）
+
+- `POST /query` - 查询重写（首轮）
+- `POST /select` - 选择候选并回答
+- `POST /reject` - 拒绝并重新生成
+- `POST /message` - 聊天消息
+
+### 通用端点
+
+- `GET /health` - 健康检查
+- `GET /version` - 版本信息
+- `GET /models` - 可用模型列表
+- `POST /set_model` - 设置模型
+
+详细 API 文档请访问：http://127.0.0.1:8010/docs
+
+## 常见问题
+
+### 1. 无法连接 Ollama 服务
+
+确保 Ollama 服务正在运行：
+```bash
+ollama serve
+```
+
+检查配置文件中的 `OLLAMA_HOST` 是否正确。
+
+### 2. 模型不可用
+
+通过 `/models` 端点查看可用模型列表，使用 `/set_model` 切换模型。
+
+### 3. 生成速度慢
+
+- 调整 `GENERATION_POOL_SIZE` 减少候选数量
+- 使用更小的模型（如 `qwen3:4b`）
+- 确保 Ollama 使用 GPU 加速
+
+### 4. 界面显示异常
+
+硬刷新浏览器缓存：
+- **Mac**: `Cmd + Shift + R`
+- **Windows/Linux**: `Ctrl + Shift + R`
+
+---
+
+<details>
+<summary><b>原始 README（点击展开）</b></summary>
+
 - 项目简介
 
 - OPRO Prompt Optimizer：面向提示优化的交互式系统，支持多轮拒选/再生成、语义聚类去重与 Top‑K 代表选择。
@@ -65,3 +280,5 @@
 - 第二轮无相关候选：使用 POST /query_from_message 基于最近消息再生候选 _qwen_xinference_demo/api.py:193-206
 - 立即回答诉求：用 POST /answer 先答后给候选 _qwen_xinference_demo/api.py:211-219
 - 端口与地址访问差异：在启动命令中明确 --host 0.0.0.0 --port 8010 ，本地浏览器建议访问 127.0.0.1
+
+</details>

_qwen_xinference_demo/api.py

@@ -2,14 +2,30 @@ from fastapi import FastAPI, HTTPException, Request
 from fastapi.responses import RedirectResponse, FileResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
 from pydantic import BaseModel
+from typing import List, Tuple, Optional
 import config
 
+# Legacy session management (query rewriting)
 from .opro.session_state import create_session, get_session, update_session_add_candidates, log_user_choice
 from .opro.session_state import log_user_reject
 from .opro.session_state import set_selected_prompt, log_chat_message
 from .opro.session_state import set_session_model
 from .opro.session_state import USER_FEEDBACK_LOG
+
+# True OPRO session management
+from .opro.session_state import (
+    create_opro_run, get_opro_run, update_opro_iteration,
+    add_opro_evaluation, get_opro_trajectory, set_opro_test_cases,
+    complete_opro_run, list_opro_runs
+)
+
+# Optimization functions
 from .opro.user_prompt_optimizer import generate_candidates
+from .opro.user_prompt_optimizer import (
+    generate_system_instruction_candidates,
+    evaluate_system_instruction
+)
+
 from .opro.ollama_client import call_qwen
 from .opro.ollama_client import list_models
 
@@ -23,8 +39,9 @@ app = FastAPI(
     openapi_tags=[
         {"name": "health", "description": "健康检查"},
         {"name": "models", "description": "模型列表与设置"},
-        {"name": "sessions", "description": "会话管理"},
+        {"name": "sessions", "description": "会话管理（旧版查询重写）"},
-        {"name": "opro", "description": "提示优化候选生成与选择/拒绝"},
+        {"name": "opro-legacy", "description": "旧版提示优化（查询重写）"},
+        {"name": "opro-true", "description": "真正的OPRO（系统指令优化）"},
         {"name": "chat", "description": "会话聊天"},
         {"name": "ui", "description": "静态页面"}
     ]
@@ -89,14 +106,69 @@ class SetModelReq(BaseModel):
     session_id: str
     model_name: str
 
-@app.post("/start", tags=["opro"])
+
+# ============================================================================
+# TRUE OPRO REQUEST MODELS
+# ============================================================================
+
+class TestCase(BaseModel):
+    """A single test case for OPRO evaluation."""
+    input: str
+    expected_output: str
+
+
+class CreateOPRORunReq(BaseModel):
+    """Request to create a new OPRO optimization run."""
+    task_description: str
+    test_cases: Optional[List[TestCase]] = None
+    model_name: Optional[str] = None
+
+
+class OPROIterateReq(BaseModel):
+    """Request to run one OPRO iteration."""
+    run_id: str
+    top_k: Optional[int] = None
+
+
+class OPROEvaluateReq(BaseModel):
+    """Request to evaluate a system instruction."""
+    run_id: str
+    instruction: str
+
+
+class OPROAddTestCasesReq(BaseModel):
+    """Request to add test cases to an OPRO run."""
+    run_id: str
+    test_cases: List[TestCase]
+
+
+class OPROGenerateAndEvaluateReq(BaseModel):
+    """Request to generate and auto-evaluate candidates (for chat-like UX)."""
+    run_id: str
+    top_k: Optional[int] = None
+    pool_size: Optional[int] = None
+    auto_evaluate: Optional[bool] = True  # If False, use diversity-based selection only
+
+
+class OPROExecuteReq(BaseModel):
+    """Request to execute a system instruction with user input."""
+    instruction: str
+    user_input: str
+    model_name: Optional[str] = None
+
+
+# ============================================================================
+# LEGACY ENDPOINTS (Query Rewriting - NOT true OPRO)
+# ============================================================================
+
+@app.post("/start", tags=["opro-legacy"])
 def start(req: StartReq):
     sid = create_session(req.query)
     cands = generate_candidates(req.query, [], model_name=get_session(sid).get("model_name"))
     update_session_add_candidates(sid, cands)
     return ok({"session_id": sid, "round": 0, "candidates": cands})
 
-@app.post("/next", tags=["opro"])
+@app.post("/next", tags=["opro-legacy"])
 def next_round(req: NextReq):
     s = get_session(req.session_id)
     if not s:
@@ -110,7 +182,7 @@ def next_round(req: NextReq):
     update_session_add_candidates(req.session_id, cands)
     return ok({"session_id": req.session_id, "round": s["round"], "candidates": cands})
 
-@app.post("/select", tags=["opro"])
+@app.post("/select", tags=["opro-legacy"])
 def select(req: SelectReq):
     s = get_session(req.session_id)
     if not s:
@@ -138,7 +210,7 @@ def select(req: SelectReq):
         pass
     return ok({"prompt": req.choice, "answer": ans})
 
-@app.post("/reject", tags=["opro"])
+@app.post("/reject", tags=["opro-legacy"])
 def reject(req: RejectReq):
     s = get_session(req.session_id)
     if not s:
@@ -151,7 +223,7 @@ class QueryReq(BaseModel):
     query: str
     session_id: str | None = None
 
-@app.post("/query", tags=["opro"])
+@app.post("/query", tags=["opro-legacy"])
 def query(req: QueryReq):
     if req.session_id:
         s = get_session(req.session_id)
@@ -240,7 +312,7 @@ def message(req: MessageReq):
 class QueryFromMsgReq(BaseModel):
     session_id: str
 
-@app.post("/query_from_message", tags=["opro"])
+@app.post("/query_from_message", tags=["opro-legacy"])
 def query_from_message(req: QueryFromMsgReq):
     s = get_session(req.session_id)
     if not s:
@@ -258,7 +330,7 @@ def query_from_message(req: QueryFromMsgReq):
 class AnswerReq(BaseModel):
     query: str
 
-@app.post("/answer", tags=["opro"])
+@app.post("/answer", tags=["opro-legacy"])
 def answer(req: AnswerReq):
     sid = create_session(req.query)
     log_chat_message(sid, "user", req.query)
@@ -282,3 +354,287 @@ def set_model(req: SetModelReq):
         raise AppException(400, f"model not available: {req.model_name}", "MODEL_NOT_AVAILABLE")
     set_session_model(req.session_id, req.model_name)
     return ok({"session_id": req.session_id, "model_name": req.model_name})
+
+
+# ============================================================================
+# TRUE OPRO ENDPOINTS (System Instruction Optimization)
+# ============================================================================
+
+@app.post("/opro/create", tags=["opro-true"])
+def opro_create_run(req: CreateOPRORunReq):
+    """
+    Create a new OPRO optimization run.
+
+    This starts a new system instruction optimization process for a given task.
+    """
+    # Convert test cases from Pydantic models to tuples
+    test_cases = None
+    if req.test_cases:
+        test_cases = [(tc.input, tc.expected_output) for tc in req.test_cases]
+
+    run_id = create_opro_run(
+        task_description=req.task_description,
+        test_cases=test_cases,
+        model_name=req.model_name
+    )
+
+    run = get_opro_run(run_id)
+
+    return ok({
+        "run_id": run_id,
+        "task_description": run["task_description"],
+        "num_test_cases": len(run["test_cases"]),
+        "iteration": run["iteration"],
+        "status": run["status"]
+    })
+
+
+@app.post("/opro/iterate", tags=["opro-true"])
+def opro_iterate(req: OPROIterateReq):
+    """
+    Run one OPRO iteration: generate new system instruction candidates.
+
+    This generates optimized system instructions based on the performance trajectory.
+    """
+    run = get_opro_run(req.run_id)
+    if not run:
+        raise AppException(404, "OPRO run not found", "RUN_NOT_FOUND")
+
+    # Get trajectory for optimization
+    trajectory = get_opro_trajectory(req.run_id)
+
+    # Generate candidates
+    top_k = req.top_k or config.TOP_K
+    try:
+        candidates = generate_system_instruction_candidates(
+            task_description=run["task_description"],
+            trajectory=trajectory if trajectory else None,
+            top_k=top_k,
+            model_name=run["model_name"]
+        )
+    except Exception as e:
+        raise AppException(500, f"Failed to generate candidates: {e}", "GENERATION_ERROR")
+
+    # Update run with new candidates
+    update_opro_iteration(req.run_id, candidates)
+
+    return ok({
+        "run_id": req.run_id,
+        "iteration": run["iteration"] + 1,
+        "candidates": candidates,
+        "num_candidates": len(candidates),
+        "best_score": run["best_score"]
+    })
+
+
+@app.post("/opro/evaluate", tags=["opro-true"])
+def opro_evaluate(req: OPROEvaluateReq):
+    """
+    Evaluate a system instruction on the test cases.
+
+    This scores the instruction and updates the performance trajectory.
+    """
+    run = get_opro_run(req.run_id)
+    if not run:
+        raise AppException(404, "OPRO run not found", "RUN_NOT_FOUND")
+
+    if not run["test_cases"]:
+        raise AppException(400, "No test cases defined for this run", "NO_TEST_CASES")
+
+    # Evaluate the instruction
+    try:
+        score = evaluate_system_instruction(
+            system_instruction=req.instruction,
+            test_cases=run["test_cases"],
+            model_name=run["model_name"]
+        )
+    except Exception as e:
+        raise AppException(500, f"Evaluation failed: {e}", "EVALUATION_ERROR")
+
+    # Add to trajectory
+    add_opro_evaluation(req.run_id, req.instruction, score)
+
+    # Get updated run info
+    run = get_opro_run(req.run_id)
+
+    return ok({
+        "run_id": req.run_id,
+        "instruction": req.instruction,
+        "score": score,
+        "best_score": run["best_score"],
+        "is_new_best": score == run["best_score"] and score > 0
+    })
+
+
+@app.get("/opro/runs", tags=["opro-true"])
+def opro_list_runs():
+    """
+    List all OPRO optimization runs.
+    """
+    runs = list_opro_runs()
+    return ok({"runs": runs, "total": len(runs)})
+
+
+@app.get("/opro/run/{run_id}", tags=["opro-true"])
+def opro_get_run(run_id: str):
+    """
+    Get detailed information about an OPRO run.
+    """
+    run = get_opro_run(run_id)
+    if not run:
+        raise AppException(404, "OPRO run not found", "RUN_NOT_FOUND")
+
+    # Get sorted trajectory
+    trajectory = get_opro_trajectory(run_id)
+
+    return ok({
+        "run_id": run_id,
+        "task_description": run["task_description"],
+        "iteration": run["iteration"],
+        "status": run["status"],
+        "best_score": run["best_score"],
+        "best_instruction": run["best_instruction"],
+        "num_test_cases": len(run["test_cases"]),
+        "test_cases": [{"input": tc[0], "expected_output": tc[1]} for tc in run["test_cases"]],
+        "trajectory": [{"instruction": inst, "score": score} for inst, score in trajectory[:10]],  # Top 10
+        "current_candidates": run["current_candidates"]
+    })
+
+
+@app.post("/opro/test_cases", tags=["opro-true"])
+def opro_add_test_cases(req: OPROAddTestCasesReq):
+    """
+    Add or update test cases for an OPRO run.
+    """
+    run = get_opro_run(req.run_id)
+    if not run:
+        raise AppException(404, "OPRO run not found", "RUN_NOT_FOUND")
+
+    # Convert test cases
+    test_cases = [(tc.input, tc.expected_output) for tc in req.test_cases]
+
+    # Update test cases
+    set_opro_test_cases(req.run_id, test_cases)
+
+    return ok({
+        "run_id": req.run_id,
+        "num_test_cases": len(test_cases),
+        "test_cases": [{"input": tc[0], "expected_output": tc[1]} for tc in test_cases]
+    })
+
+
+@app.post("/opro/generate_and_evaluate", tags=["opro-true"])
+def opro_generate_and_evaluate(req: OPROGenerateAndEvaluateReq):
+    """
+    Generate candidates and auto-evaluate them (for chat-like UX).
+
+    This is the main endpoint for the chat interface. It:
+    1. Generates candidates based on trajectory
+    2. Auto-evaluates them (if test cases exist and auto_evaluate=True)
+    3. Returns top-k sorted by score (or diversity if no evaluation)
+    """
+    run = get_opro_run(req.run_id)
+    if not run:
+        raise AppException(404, "OPRO run not found", "RUN_NOT_FOUND")
+
+    top_k = req.top_k or config.TOP_K
+    pool_size = req.pool_size or config.GENERATION_POOL_SIZE
+
+    # Get trajectory for optimization
+    trajectory = get_opro_trajectory(req.run_id)
+
+    # Generate candidates
+    try:
+        candidates = generate_system_instruction_candidates(
+            task_description=run["task_description"],
+            trajectory=trajectory if trajectory else None,
+            top_k=pool_size,  # Generate pool_size candidates first
+            pool_size=pool_size,
+            model_name=run["model_name"]
+        )
+    except Exception as e:
+        raise AppException(500, f"Failed to generate candidates: {e}", "GENERATION_ERROR")
+
+    # Decide whether to evaluate
+    should_evaluate = req.auto_evaluate and len(run["test_cases"]) > 0
+
+    if should_evaluate:
+        # Auto-evaluate all candidates
+        scored_candidates = []
+        for candidate in candidates:
+            try:
+                score = evaluate_system_instruction(
+                    system_instruction=candidate,
+                    test_cases=run["test_cases"],
+                    model_name=run["model_name"]
+                )
+                scored_candidates.append({"instruction": candidate, "score": score})
+
+                # Add to trajectory
+                add_opro_evaluation(req.run_id, candidate, score)
+            except Exception as e:
+                # If evaluation fails, assign score 0
+                scored_candidates.append({"instruction": candidate, "score": 0.0})
+
+        # Sort by score (highest first)
+        scored_candidates.sort(key=lambda x: x["score"], reverse=True)
+
+        # Return top-k
+        top_candidates = scored_candidates[:top_k]
+
+        # Update iteration
+        update_opro_iteration(req.run_id, [c["instruction"] for c in top_candidates])
+
+        return ok({
+            "run_id": req.run_id,
+            "candidates": top_candidates,
+            "iteration": run["iteration"] + 1,
+            "evaluated": True,
+            "best_score": run["best_score"]
+        })
+    else:
+        # No evaluation - use diversity-based selection (already done by clustering)
+        # Just return the candidates without scores
+        top_candidates = [
+            {"instruction": candidate, "score": None}
+            for candidate in candidates[:top_k]
+        ]
+
+        # Update iteration
+        update_opro_iteration(req.run_id, [c["instruction"] for c in top_candidates])
+
+        return ok({
+            "run_id": req.run_id,
+            "candidates": top_candidates,
+            "iteration": run["iteration"] + 1,
+            "evaluated": False,
+            "best_score": run["best_score"]
+        })
+
+
+@app.post("/opro/execute", tags=["opro-true"])
+def opro_execute(req: OPROExecuteReq):
+    """
+    Execute a system instruction with user input.
+
+    This uses the selected instruction as a system prompt and calls the LLM.
+    """
+    try:
+        # Construct full prompt with system instruction
+        full_prompt = f"{req.instruction}\n\n{req.user_input}"
+
+        # Call LLM
+        response = call_qwen(
+            full_prompt,
+            temperature=0.2,
+            max_tokens=1024,
+            model_name=req.model_name
+        )
+
+        return ok({
+            "instruction": req.instruction,
+            "user_input": req.user_input,
+            "response": response
+        })
+    except Exception as e:
+        raise AppException(500, f"Execution failed: {e}", "EXECUTION_ERROR")

_qwen_xinference_demo/opro/prompt_utils.py

@@ -1,4 +1,14 @@
+from typing import List, Tuple
+
+# ============================================================================
+# OLD FUNCTIONS (Query Rewriting - NOT true OPRO, kept for compatibility)
+# ============================================================================
+
 def refine_instruction(query: str) -> str:
+    """
+    LEGACY: Generates query rewrites (NOT true OPRO).
+    This is query expansion, not system instruction optimization.
+    """
     return f"""
 你是一个“问题澄清与重写助手”。
 请根据用户的原始问题：
@@ -7,6 +17,9 @@ def refine_instruction(query: str) -> str:
 """
 
 def refine_instruction_with_history(query: str, rejected_list: list) -> str:
+    """
+    LEGACY: Generates query rewrites with rejection history (NOT true OPRO).
+    """
     rejected_text = "\n".join(f"- {r}" for r in rejected_list) if rejected_list else ""
     return f"""
 你是一个“问题澄清与重写助手”。
@@ -18,3 +31,100 @@ def refine_instruction_with_history(query: str, rejected_list: list) -> str:
 
 请从新的角度重新生成至少20条不同的改写问题，每条单独一行。
 """
+
+
+# ============================================================================
+# TRUE OPRO FUNCTIONS (System Instruction Optimization)
+# ============================================================================
+
+def generate_initial_system_instruction_candidates(task_description: str, pool_size: int = None) -> str:
+    """
+    TRUE OPRO: Generates initial candidate System Instructions for a new OPRO run.
+
+    Args:
+        task_description: Description of the task the LLM should perform
+        pool_size: Number of candidates to generate (defaults to config.GENERATION_POOL_SIZE)
+
+    Returns:
+        Meta-prompt that instructs the optimizer LLM to generate system instruction candidates
+    """
+    import config
+    pool_size = pool_size or config.GENERATION_POOL_SIZE
+
+    return f"""
+你是一个"系统指令生成助手"。
+目标任务描述：
+【{task_description}】
+
+请根据以上任务，生成 {pool_size} 条高质量、风格各异的"System Instruction"候选指令。
+
+要求：
+1. 每条指令必须有明显不同的风格和侧重点
+2. 覆盖不同的实现策略（例如：简洁型、详细型、示例型、角色扮演型、步骤型等）
+3. 这些指令应指导LLM的行为和输出格式，以最大化任务性能
+4. 每条指令单独成行，不包含编号或额外说明
+5. 所有生成的指令必须使用简体中文
+
+生成 {pool_size} 条指令：
+"""
+
+
+def generate_optimized_system_instruction(
+    task_description: str,
+    trajectory: List[Tuple[str, float]],
+    pool_size: int = None
+) -> str:
+    """
+    TRUE OPRO: Analyzes performance trajectory and generates optimized System Instructions.
+
+    This is the core OPRO function that uses an LLM as an optimizer to improve
+    system instructions based on historical performance scores.
+
+    Args:
+        task_description: Description of the task the LLM should perform
+        trajectory: List of (instruction, score) tuples, sorted by score (highest first)
+        pool_size: Number of candidates to generate (defaults to config.GENERATION_POOL_SIZE)
+
+    Returns:
+        Meta-prompt that instructs the optimizer LLM to generate better system instructions
+    """
+    import config
+    pool_size = pool_size or config.GENERATION_POOL_SIZE
+
+    if not trajectory:
+        # If no trajectory, fall back to initial generation
+        return generate_initial_system_instruction_candidates(task_description, pool_size)
+
+    # Format the trajectory for the Optimizer LLM
+    formatted_history = "\n".join(
+        f"--- Instruction Score: {score:.4f}\n{instruction}"
+        for instruction, score in trajectory
+    )
+
+    # Determine the current highest score to set the optimization goal
+    highest_score = max(score for _, score in trajectory)
+
+    # Construct the Meta-Prompt (The OPRO Instruction)
+    return f"""
+你是一个"System Prompt 优化器"。
+你的任务是改进一个LLM的系统指令，以最大化其在以下任务中的性能：
+【{task_description}】
+
+---
+**历史性能轨迹 (Instructions and Scores):**
+{formatted_history}
+---
+**当前最高得分: {highest_score:.4f}**
+
+请分析得分最高的指令的特点和得分最低指令的缺陷。
+然后，生成 {pool_size} 条新的、有潜力超越 {highest_score:.4f} 分的System Instruction。
+
+要求：
+1. 每条指令必须有明显不同的改进策略
+2. 结合高分指令的优点，避免低分指令的缺陷
+3. 探索新的优化方向和表达方式
+4. 每条指令单独成行，不包含编号或额外说明
+5. 所有生成的指令必须使用简体中文
+
+生成 {pool_size} 条优化后的指令：
+"""

_qwen_xinference_demo/opro/session_state.py

@@ -1,8 +1,14 @@
 import uuid
+from typing import List, Tuple, Dict, Any
 
+# Legacy session storage (for query rewriting)
 SESSIONS = {}
 USER_FEEDBACK_LOG = []
 
+# OPRO session storage (for system instruction optimization)
+OPRO_RUNS = {}
+OPRO_RUN_LOG = []
+
 def create_session(query: str) -> str:
     sid = uuid.uuid4().hex
     SESSIONS[sid] = {
@@ -54,3 +60,167 @@ def set_session_model(sid: str, model_name: str | None):
     s = SESSIONS.get(sid)
     if s is not None:
         s["model_name"] = model_name
+
+
+# ============================================================================
+# TRUE OPRO SESSION MANAGEMENT
+# ============================================================================
+
+def create_opro_run(
+    task_description: str,
+    test_cases: List[Tuple[str, str]] = None,
+    model_name: str = None
+) -> str:
+    """
+    Create a new OPRO optimization run.
+
+    Args:
+        task_description: Description of the task to optimize for
+        test_cases: List of (input, expected_output) tuples for evaluation
+        model_name: Optional model name to use
+
+    Returns:
+        run_id: Unique identifier for this OPRO run
+    """
+    run_id = uuid.uuid4().hex
+    OPRO_RUNS[run_id] = {
+        "task_description": task_description,
+        "test_cases": test_cases or [],
+        "model_name": model_name,
+        "iteration": 0,
+        "trajectory": [],  # List of (instruction, score) tuples
+        "best_instruction": None,
+        "best_score": 0.0,
+        "current_candidates": [],
+        "created_at": uuid.uuid1().time,
+        "status": "active"  # active, completed, failed
+    }
+    return run_id
+
+
+def get_opro_run(run_id: str) -> Dict[str, Any]:
+    """Get OPRO run by ID."""
+    return OPRO_RUNS.get(run_id)
+
+
+def update_opro_iteration(
+    run_id: str,
+    candidates: List[str],
+    scores: List[float] = None
+):
+    """
+    Update OPRO run with new iteration results.
+
+    Args:
+        run_id: OPRO run identifier
+        candidates: List of system instruction candidates
+        scores: Optional list of scores (if evaluated)
+    """
+    run = OPRO_RUNS.get(run_id)
+    if not run:
+        return
+
+    run["iteration"] += 1
+    run["current_candidates"] = candidates
+
+    # If scores provided, update trajectory
+    if scores and len(scores) == len(candidates):
+        for candidate, score in zip(candidates, scores):
+            run["trajectory"].append((candidate, score))
+
+            # Update best if this is better
+            if score > run["best_score"]:
+                run["best_score"] = score
+                run["best_instruction"] = candidate
+
+    # Log the iteration
+    OPRO_RUN_LOG.append({
+        "run_id": run_id,
+        "iteration": run["iteration"],
+        "num_candidates": len(candidates),
+        "best_score": run["best_score"]
+    })
+
+
+def add_opro_evaluation(
+    run_id: str,
+    instruction: str,
+    score: float
+):
+    """
+    Add a single evaluation result to OPRO run.
+
+    Args:
+        run_id: OPRO run identifier
+        instruction: System instruction that was evaluated
+        score: Performance score
+    """
+    run = OPRO_RUNS.get(run_id)
+    if not run:
+        return
+
+    # Add to trajectory
+    run["trajectory"].append((instruction, score))
+
+    # Update best if this is better
+    if score > run["best_score"]:
+        run["best_score"] = score
+        run["best_instruction"] = instruction
+
+
+def get_opro_trajectory(run_id: str) -> List[Tuple[str, float]]:
+    """
+    Get the performance trajectory for an OPRO run.
+
+    Returns:
+        List of (instruction, score) tuples sorted by score (highest first)
+    """
+    run = OPRO_RUNS.get(run_id)
+    if not run:
+        return []
+
+    trajectory = run["trajectory"]
+    return sorted(trajectory, key=lambda x: x[1], reverse=True)
+
+
+def set_opro_test_cases(
+    run_id: str,
+    test_cases: List[Tuple[str, str]]
+):
+    """
+    Set or update test cases for an OPRO run.
+
+    Args:
+        run_id: OPRO run identifier
+        test_cases: List of (input, expected_output) tuples
+    """
+    run = OPRO_RUNS.get(run_id)
+    if run:
+        run["test_cases"] = test_cases
+
+
+def complete_opro_run(run_id: str):
+    """Mark an OPRO run as completed."""
+    run = OPRO_RUNS.get(run_id)
+    if run:
+        run["status"] = "completed"
+
+
+def list_opro_runs() -> List[Dict[str, Any]]:
+    """
+    List all OPRO runs with summary information.
+
+    Returns:
+        List of run summaries
+    """
+    return [
+        {
+            "run_id": run_id,
+            "task_description": run["task_description"][:100] + "..." if len(run["task_description"]) > 100 else run["task_description"],
+            "iteration": run["iteration"],
+            "best_score": run["best_score"],
+            "num_test_cases": len(run["test_cases"]),
+            "status": run["status"]
+        }
+        for run_id, run in OPRO_RUNS.items()
+    ]

_qwen_xinference_demo/opro/user_prompt_optimizer.py

@@ -1,12 +1,18 @@
 import re
 import numpy as np
+from typing import List, Tuple
 from sklearn.cluster import AgglomerativeClustering
 from sklearn.metrics.pairwise import cosine_similarity
 import config
 
 from .ollama_client import call_qwen
 from .xinference_client import embed_texts
-from .prompt_utils import refine_instruction, refine_instruction_with_history
+from .prompt_utils import (
+    refine_instruction,
+    refine_instruction_with_history,
+    generate_initial_system_instruction_candidates,
+    generate_optimized_system_instruction
+)
 
 def parse_candidates(raw: str) -> list:
     lines = [l.strip() for l in re.split(r'\r?\n', raw) if l.strip()]
@@ -44,6 +50,10 @@ def cluster_and_select(candidates: list, top_k=config.TOP_K, distance_threshold=
     return selected[:top_k]
 
 def generate_candidates(query: str, rejected=None, top_k=config.TOP_K, model_name=None):
+    """
+    LEGACY: Query rewriting function (NOT true OPRO).
+    Kept for backward compatibility with existing API endpoints.
+    """
     rejected = rejected or []
     if rejected:
         prompt = refine_instruction_with_history(query, rejected)
@@ -53,3 +63,87 @@ def generate_candidates(query: str, rejected=None, top_k=config.TOP_K, model_nam
     raw = call_qwen(prompt, temperature=0.9, max_tokens=1024, model_name=model_name)
     all_candidates = parse_candidates(raw)
     return cluster_and_select(all_candidates, top_k=top_k)
+
+
+# ============================================================================
+# TRUE OPRO FUNCTIONS (System Instruction Optimization)
+# ============================================================================
+
+def generate_system_instruction_candidates(
+    task_description: str,
+    trajectory: List[Tuple[str, float]] = None,
+    top_k: int = config.TOP_K,
+    pool_size: int = None,
+    model_name: str = None
+) -> List[str]:
+    """
+    TRUE OPRO: Generates optimized system instruction candidates.
+
+    This is the core OPRO function that generates system instructions based on
+    performance trajectory (if available) or initial candidates (if starting fresh).
+
+    Args:
+        task_description: Description of the task the LLM should perform
+        trajectory: Optional list of (instruction, score) tuples from previous iterations
+        top_k: Number of diverse candidates to return (default: config.TOP_K = 5)
+        pool_size: Number of candidates to generate before clustering (default: config.GENERATION_POOL_SIZE = 10)
+        model_name: Optional model name to use for generation
+
+    Returns:
+        List of top-k diverse system instruction candidates
+    """
+    pool_size = pool_size or config.GENERATION_POOL_SIZE
+
+    # Generate the meta-prompt based on whether we have trajectory data
+    if trajectory and len(trajectory) > 0:
+        # Sort trajectory by score (highest first)
+        sorted_trajectory = sorted(trajectory, key=lambda x: x[1], reverse=True)
+        meta_prompt = generate_optimized_system_instruction(task_description, sorted_trajectory, pool_size)
+    else:
+        # No trajectory yet, generate initial candidates
+        meta_prompt = generate_initial_system_instruction_candidates(task_description, pool_size)
+
+    # Use the optimizer LLM to generate candidates
+    raw = call_qwen(meta_prompt, temperature=0.9, max_tokens=1024, model_name=model_name)
+
+    # Parse the generated candidates
+    all_candidates = parse_candidates(raw)
+
+    # Cluster and select diverse representatives
+    return cluster_and_select(all_candidates, top_k=top_k)
+
+
+def evaluate_system_instruction(
+    system_instruction: str,
+    test_cases: List[Tuple[str, str]],
+    model_name: str = None
+) -> float:
+    """
+    TRUE OPRO: Evaluates a system instruction's performance on test cases.
+
+    Args:
+        system_instruction: The system instruction to evaluate
+        test_cases: List of (input, expected_output) tuples
+        model_name: Optional model name to use for evaluation
+
+    Returns:
+        Performance score (0.0 to 1.0)
+    """
+    if not test_cases:
+        return 0.0
+
+    correct = 0
+    total = len(test_cases)
+
+    for input_text, expected_output in test_cases:
+        # Construct the full prompt with system instruction
+        full_prompt = f"{system_instruction}\n\n{input_text}"
+
+        # Get LLM response
+        response = call_qwen(full_prompt, temperature=0.2, max_tokens=512, model_name=model_name)
+
+        # Simple exact match scoring (can be replaced with more sophisticated metrics)
+        if expected_output.strip().lower() in response.strip().lower():
+            correct += 1
+
+    return correct / total

config.py

@@ -14,6 +14,7 @@ DEFAULT_EMBED_MODEL = "qwen3-embedding:4b"
 XINFERENCE_EMBED_URL = "http://127.0.0.1:9997/models/bge-base-zh/embed"
 
 # Clustering/selection
-TOP_K = 5
+GENERATION_POOL_SIZE = 10  # Generate this many candidates before clustering
+TOP_K = 5                   # Return this many diverse candidates to user
 CLUSTER_DISTANCE_THRESHOLD = 0.15
 

examples/opro_demo.py

@@ -0,0 +1,164 @@
+"""
+TRUE OPRO Demo Script
+
+This script demonstrates the true OPRO (Optimization by PROmpting) functionality.
+It shows how to:
+1. Generate initial system instruction candidates
+2. Evaluate them on test cases
+3. Use the performance trajectory to generate better candidates
+"""
+
+import sys
+sys.path.insert(0, '.')
+
+from _qwen_xinference_demo.opro.user_prompt_optimizer import (
+    generate_system_instruction_candidates,
+    evaluate_system_instruction
+)
+import config
+
+
+def demo_opro_workflow():
+    """
+    Demonstrates a complete OPRO optimization workflow.
+    """
+    print("=" * 80)
+    print("TRUE OPRO Demo - System Instruction Optimization")
+    print("=" * 80)
+    print(f"Pool Size: {config.GENERATION_POOL_SIZE} candidates → Clustered to Top {config.TOP_K}")
+
+    # Define the task
+    task_description = """
+任务：将用户输入的中文句子翻译成英文。
+要求：翻译准确、自然、符合英语表达习惯。
+"""
+
+    print(f"\n📋 Task Description:\n{task_description}")
+    
+    # Define test cases for evaluation
+    test_cases = [
+        ("你好，很高兴见到你", "Hello, nice to meet you"),
+        ("今天天气真好", "The weather is really nice today"),
+        ("我喜欢学习编程", "I like learning programming"),
+        ("这本书很有趣", "This book is very interesting"),
+    ]
+    
+    print(f"\n🧪 Test Cases: {len(test_cases)} examples")
+    for i, (input_text, expected) in enumerate(test_cases, 1):
+        print(f"  {i}. '{input_text}' → '{expected}'")
+    
+    # Iteration 1: Generate initial candidates
+    print("\n" + "=" * 80)
+    print("🔄 Iteration 1: Generating Initial System Instruction Candidates")
+    print("=" * 80)
+    
+    print("\n⏳ Generating candidates... (this may take a moment)")
+    candidates_round1 = generate_system_instruction_candidates(
+        task_description=task_description,
+        trajectory=None,  # No history yet
+        top_k=3,
+        model_name=None  # Use default model
+    )
+    
+    print(f"\n✅ Generated {len(candidates_round1)} candidates:")
+    for i, candidate in enumerate(candidates_round1, 1):
+        print(f"\n  Candidate {i}:")
+        print(f"  {candidate[:100]}..." if len(candidate) > 100 else f"  {candidate}")
+    
+    # Evaluate each candidate
+    print("\n" + "-" * 80)
+    print("📊 Evaluating Candidates on Test Cases")
+    print("-" * 80)
+    
+    trajectory = []
+    for i, candidate in enumerate(candidates_round1, 1):
+        print(f"\n⏳ Evaluating Candidate {i}...")
+        score = evaluate_system_instruction(
+            system_instruction=candidate,
+            test_cases=test_cases,
+            model_name=None
+        )
+        trajectory.append((candidate, score))
+        print(f"  Score: {score:.2%}")
+    
+    # Sort by score
+    trajectory.sort(key=lambda x: x[1], reverse=True)
+    
+    print("\n📈 Performance Summary (Round 1):")
+    for i, (candidate, score) in enumerate(trajectory, 1):
+        print(f"  {i}. Score: {score:.2%} - {candidate[:60]}...")
+    
+    best_score = trajectory[0][1]
+    print(f"\n🏆 Best Score: {best_score:.2%}")
+    
+    # Iteration 2: Generate optimized candidates based on trajectory
+    print("\n" + "=" * 80)
+    print("🔄 Iteration 2: Generating Optimized System Instructions")
+    print("=" * 80)
+    print(f"\n💡 Using performance trajectory to generate better candidates...")
+    print(f"   Goal: Beat current best score of {best_score:.2%}")
+    
+    print("\n⏳ Generating optimized candidates...")
+    candidates_round2 = generate_system_instruction_candidates(
+        task_description=task_description,
+        trajectory=trajectory,  # Use performance history
+        top_k=3,
+        model_name=None
+    )
+    
+    print(f"\n✅ Generated {len(candidates_round2)} optimized candidates:")
+    for i, candidate in enumerate(candidates_round2, 1):
+        print(f"\n  Candidate {i}:")
+        print(f"  {candidate[:100]}..." if len(candidate) > 100 else f"  {candidate}")
+    
+    # Evaluate new candidates
+    print("\n" + "-" * 80)
+    print("📊 Evaluating Optimized Candidates")
+    print("-" * 80)
+    
+    for i, candidate in enumerate(candidates_round2, 1):
+        print(f"\n⏳ Evaluating Optimized Candidate {i}...")
+        score = evaluate_system_instruction(
+            system_instruction=candidate,
+            test_cases=test_cases,
+            model_name=None
+        )
+        trajectory.append((candidate, score))
+        print(f"  Score: {score:.2%}")
+        if score > best_score:
+            print(f"  🎉 NEW BEST! Improved from {best_score:.2%} to {score:.2%}")
+            best_score = score
+    
+    # Final summary
+    trajectory.sort(key=lambda x: x[1], reverse=True)
+    
+    print("\n" + "=" * 80)
+    print("🏁 Final Results")
+    print("=" * 80)
+    print(f"\n🏆 Best System Instruction (Score: {trajectory[0][1]:.2%}):")
+    print(f"\n{trajectory[0][0]}")
+    
+    print("\n📊 All Candidates Ranked:")
+    for i, (candidate, score) in enumerate(trajectory[:5], 1):
+        print(f"\n  {i}. Score: {score:.2%}")
+        print(f"     {candidate[:80]}...")
+    
+    print("\n" + "=" * 80)
+    print("✅ OPRO Demo Complete!")
+    print("=" * 80)
+
+
+if __name__ == "__main__":
+    print("\n⚠️  NOTE: This demo requires:")
+    print("   1. Ollama running locally (http://127.0.0.1:11434)")
+    print("   2. A Qwen model available (e.g., qwen3:8b)")
+    print("   3. An embedding model (e.g., qwen3-embedding:4b)")
+    print("\n   Press Ctrl+C to cancel, or Enter to continue...")
+    
+    try:
+        input()
+        demo_opro_workflow()
+    except KeyboardInterrupt:
+        print("\n\n❌ Demo cancelled by user.")
+        sys.exit(0)
+

frontend/opro.html

@@ -0,0 +1,507 @@
+<!DOCTYPE html>
+<html lang="zh-CN">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate">
+    <meta http-equiv="Pragma" content="no-cache">
+    <meta http-equiv="Expires" content="0">
+    <title>OPRO - System Instruction Optimizer</title>
+    <script crossorigin src="https://unpkg.com/react@18/umd/react.production.min.js"></script>
+    <script crossorigin src="https://unpkg.com/react-dom@18/umd/react-dom.production.min.js"></script>
+    <script src="https://cdn.tailwindcss.com"></script>
+    <style>
+        body {
+            margin: 0;
+            font-family: 'Google Sans', 'Segoe UI', Roboto, sans-serif;
+            background: #f8f9fa;
+        }
+        .chat-container { height: 100vh; display: flex; }
+        .scrollbar-hide::-webkit-scrollbar { display: none; }
+        .scrollbar-hide { -ms-overflow-style: none; scrollbar-width: none; }
+        .sidebar-collapsed { width: 60px; }
+        .sidebar-expanded { width: 260px; }
+        .instruction-card {
+            transition: all 0.15s ease;
+            border: 1px solid #e8eaed;
+        }
+        .instruction-card:hover {
+            border-color: #dadce0;
+            box-shadow: 0 1px 3px rgba(60,64,67,0.15);
+        }
+        .loading-dots::after {
+            content: '...';
+            animation: dots 1.5s steps(4, end) infinite;
+        }
+        @keyframes dots {
+            0%, 20% { content: '.'; }
+            40% { content: '..'; }
+            60%, 100% { content: '...'; }
+        }
+    </style>
+</head>
+<body>
+    <div id="root"></div>
+    
+    <script>
+        const { useState, useEffect, useRef } = React;
+        const API_BASE = 'http://127.0.0.1:8010';
+        
+        // Main App Component
+        function App() {
+            const [sidebarOpen, setSidebarOpen] = useState(false);
+            const [runs, setRuns] = useState([]);
+            const [currentRunId, setCurrentRunId] = useState(null);
+            const [messages, setMessages] = useState([]);
+            const [inputValue, setInputValue] = useState('');
+            const [loading, setLoading] = useState(false);
+            const [models, setModels] = useState([]);
+            const [selectedModel, setSelectedModel] = useState('');
+            const chatEndRef = useRef(null);
+
+            // Load runs and models on mount
+            useEffect(() => {
+                loadRuns();
+                loadModels();
+            }, []);
+
+            async function loadModels() {
+                try {
+                    const res = await fetch(`${API_BASE}/models`);
+                    const data = await res.json();
+                    if (data.success && data.data.models) {
+                        setModels(data.data.models);
+                        if (data.data.models.length > 0) {
+                            setSelectedModel(data.data.models[0]);
+                        }
+                    }
+                } catch (err) {
+                    console.error('Failed to load models:', err);
+                }
+            }
+            
+            // Auto-scroll chat
+            useEffect(() => {
+                chatEndRef.current?.scrollIntoView({ behavior: 'smooth' });
+            }, [messages]);
+            
+            async function loadRuns() {
+                try {
+                    const res = await fetch(`${API_BASE}/opro/runs`);
+                    const data = await res.json();
+                    if (data.success) {
+                        setRuns(data.data.runs || []);
+                    }
+                } catch (err) {
+                    console.error('Failed to load runs:', err);
+                }
+            }
+            
+            async function createNewRun(taskDescription) {
+                setLoading(true);
+                try {
+                    // Create run
+                    const res = await fetch(`${API_BASE}/opro/create`, {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({
+                            task_description: taskDescription,
+                            test_cases: [],
+                            model_name: selectedModel || undefined
+                        })
+                    });
+                    const data = await res.json();
+                    
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to create run');
+                    }
+                    
+                    const runId = data.data.run_id;
+                    setCurrentRunId(runId);
+                    
+                    // Add user message
+                    setMessages([{ role: 'user', content: taskDescription }]);
+                    
+                    // Generate and evaluate candidates
+                    await generateCandidates(runId);
+                    
+                    // Reload runs list
+                    await loadRuns();
+                } catch (err) {
+                    alert('创建任务失败: ' + err.message);
+                } finally {
+                    setLoading(false);
+                }
+            }
+            
+            async function generateCandidates(runId) {
+                setLoading(true);
+                try {
+                    const res = await fetch(`${API_BASE}/opro/generate_and_evaluate`, {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({
+                            run_id: runId,
+                            top_k: 5,
+                            auto_evaluate: false  // Use diversity-based selection
+                        })
+                    });
+                    const data = await res.json();
+                    
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to generate candidates');
+                    }
+                    
+                    // Add assistant message with candidates
+                    setMessages(prev => [...prev, {
+                        role: 'assistant',
+                        type: 'candidates',
+                        candidates: data.data.candidates,
+                        iteration: data.data.iteration
+                    }]);
+                } catch (err) {
+                    alert('生成候选指令失败: ' + err.message);
+                } finally {
+                    setLoading(false);
+                }
+            }
+
+            async function executeInstruction(instruction, userInput) {
+                setLoading(true);
+                try {
+                    const res = await fetch(`${API_BASE}/opro/execute`, {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({
+                            instruction: instruction,
+                            user_input: userInput || '请执行任务',
+                            model_name: selectedModel || undefined
+                        })
+                    });
+                    const data = await res.json();
+
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to execute');
+                    }
+
+                    // Add execution result
+                    setMessages(prev => [...prev, {
+                        role: 'assistant',
+                        type: 'execution',
+                        instruction: instruction,
+                        response: data.data.response
+                    }]);
+                } catch (err) {
+                    alert('执行失败: ' + err.message);
+                } finally {
+                    setLoading(false);
+                }
+            }
+
+            function handleSendMessage() {
+                const msg = inputValue.trim();
+                if (!msg || loading) return;
+
+                setInputValue('');
+
+                if (!currentRunId) {
+                    // Create new run with task description
+                    createNewRun(msg);
+                } else {
+                    // Continue optimization or execute
+                    // For now, just show message
+                    setMessages(prev => [...prev, { role: 'user', content: msg }]);
+                }
+            }
+
+            function handleContinueOptimize() {
+                if (!currentRunId || loading) return;
+                generateCandidates(currentRunId);
+            }
+
+            function handleExecute(instruction) {
+                if (loading) return;
+                const userInput = prompt('请输入要处理的内容（可选）:');
+                executeInstruction(instruction, userInput);
+            }
+
+            function handleCopyInstruction(instruction) {
+                navigator.clipboard.writeText(instruction).then(() => {
+                    // Could add a toast notification here
+                    console.log('Instruction copied to clipboard');
+                }).catch(err => {
+                    console.error('Failed to copy:', err);
+                });
+            }
+
+            function handleNewTask() {
+                setCurrentRunId(null);
+                setMessages([]);
+                setInputValue('');
+            }
+
+            async function loadRun(runId) {
+                setLoading(true);
+                try {
+                    const res = await fetch(`${API_BASE}/opro/run/${runId}`);
+                    const data = await res.json();
+
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to load run');
+                    }
+
+                    const run = data.data;
+                    setCurrentRunId(runId);
+
+                    // Reconstruct messages from run data
+                    const msgs = [
+                        { role: 'user', content: run.task_description }
+                    ];
+
+                    if (run.current_candidates && run.current_candidates.length > 0) {
+                        msgs.push({
+                            role: 'assistant',
+                            type: 'candidates',
+                            candidates: run.current_candidates.map(c => ({ instruction: c, score: null })),
+                            iteration: run.iteration
+                        });
+                    }
+
+                    setMessages(msgs);
+                } catch (err) {
+                    alert('加载任务失败: ' + err.message);
+                } finally {
+                    setLoading(false);
+                }
+            }
+
+            return React.createElement('div', { className: 'chat-container' },
+                // Sidebar
+                React.createElement('div', {
+                    className: `bg-white border-r border-gray-200 transition-all duration-300 flex flex-col ${sidebarOpen ? 'sidebar-expanded' : 'sidebar-collapsed'}`
+                },
+                    // Header area - Collapse button only
+                    React.createElement('div', { className: 'p-3 border-b border-gray-200 flex items-center justify-between' },
+                        sidebarOpen ? React.createElement('button', {
+                            onClick: () => setSidebarOpen(false),
+                            className: 'p-2 text-gray-600 hover:bg-gray-100 rounded-lg transition-colors'
+                        },
+                            React.createElement('svg', { width: '20', height: '20', viewBox: '0 0 24 24', fill: 'none', stroke: 'currentColor', strokeWidth: '2' },
+                                React.createElement('path', { d: 'M15 18l-6-6 6-6' })
+                            )
+                        ) : React.createElement('button', {
+                            onClick: () => setSidebarOpen(true),
+                            className: 'w-full p-2 text-gray-600 hover:bg-gray-100 rounded-lg transition-colors flex items-center justify-center'
+                        },
+                            React.createElement('svg', { width: '20', height: '20', viewBox: '0 0 24 24', fill: 'none', stroke: 'currentColor', strokeWidth: '2' },
+                                React.createElement('path', { d: 'M3 12h18M3 6h18M3 18h18' })
+                            )
+                        )
+                    ),
+                    // Content area
+                    React.createElement('div', { className: 'flex-1 overflow-y-auto scrollbar-hide p-2 flex flex-col' },
+                        sidebarOpen ? React.createElement(React.Fragment, null,
+                            // New task button (expanded)
+                            React.createElement('button', {
+                                onClick: handleNewTask,
+                                className: 'mb-3 px-4 py-2.5 bg-white border border-gray-300 hover:bg-gray-50 rounded-lg transition-colors flex items-center justify-center gap-2 text-gray-700 font-medium'
+                            },
+                                React.createElement('span', { className: 'text-lg' }, '+'),
+                                React.createElement('span', null, '新建会话')
+                            ),
+                            // Sessions list
+                            runs.length > 0 && React.createElement('div', { className: 'text-xs text-gray-500 mb-2 px-2' }, '会话列表'),
+                            runs.map(run =>
+                                React.createElement('div', {
+                                    key: run.run_id,
+                                    onClick: () => loadRun(run.run_id),
+                                    className: `p-3 mb-1 rounded-lg cursor-pointer transition-colors flex items-center gap-2 ${
+                                        currentRunId === run.run_id ? 'bg-gray-100' : 'hover:bg-gray-50'
+                                    }`
+                                },
+                                    React.createElement('svg', {
+                                        width: '16',
+                                        height: '16',
+                                        viewBox: '0 0 24 24',
+                                        fill: 'none',
+                                        stroke: 'currentColor',
+                                        strokeWidth: '2',
+                                        className: 'flex-shrink-0 text-gray-500'
+                                    },
+                                        React.createElement('path', { d: 'M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z' })
+                                    ),
+                                    React.createElement('div', { className: 'text-sm text-gray-800 truncate flex-1' },
+                                        run.task_description
+                                    )
+                                )
+                            )
+                        ) : React.createElement('button', {
+                            onClick: handleNewTask,
+                            className: 'p-2 text-gray-600 hover:bg-gray-100 rounded-lg transition-colors flex items-center justify-center',
+                            title: '新建会话'
+                        },
+                            React.createElement('svg', { width: '24', height: '24', viewBox: '0 0 24 24', fill: 'none', stroke: 'currentColor', strokeWidth: '2' },
+                                React.createElement('path', { d: 'M12 5v14M5 12h14' })
+                            )
+                        )
+                    )
+                ),
+
+                // Main Chat Area
+                React.createElement('div', { className: 'flex-1 flex flex-col bg-white' },
+                    // Header
+                    React.createElement('div', { className: 'px-4 py-3 border-b border-gray-200 bg-white flex items-center gap-3' },
+                        React.createElement('h1', { className: 'text-lg font-normal text-gray-800' },
+                            'OPRO'
+                        )
+                    ),
+
+                    // Chat Messages
+                    React.createElement('div', { className: 'flex-1 overflow-y-auto scrollbar-hide p-6 space-y-6 max-w-4xl mx-auto w-full' },
+                        messages.map((msg, idx) => {
+                            if (msg.role === 'user') {
+                                return React.createElement('div', { key: idx, className: 'flex justify-end' },
+                                    React.createElement('div', { className: 'max-w-2xl bg-gray-100 text-gray-800 rounded-2xl px-5 py-3' },
+                                        msg.content
+                                    )
+                                );
+                            } else if (msg.type === 'candidates') {
+                                return React.createElement('div', { key: idx, className: 'flex justify-start' },
+                                    React.createElement('div', { className: 'w-full' },
+                                        React.createElement('div', { className: 'mb-3' },
+                                            React.createElement('div', { className: 'text-sm text-gray-600' },
+                                                `优化后的提示词（第 ${msg.iteration} 轮）`
+                                            ),
+                                        ),
+                                        msg.candidates.map((cand, cidx) =>
+                                            React.createElement('div', {
+                                                key: cidx,
+                                                className: 'instruction-card bg-white rounded-xl p-5 mb-3'
+                                            },
+                                                React.createElement('div', { className: 'flex items-start gap-3' },
+                                                    React.createElement('div', { className: 'flex-shrink-0 w-7 h-7 bg-gray-200 text-gray-700 rounded-full flex items-center justify-center text-sm font-medium' },
+                                                        cidx + 1
+                                                    ),
+                                                    React.createElement('div', { className: 'flex-1' },
+                                                        React.createElement('div', { className: 'text-gray-800 mb-4 whitespace-pre-wrap leading-relaxed' },
+                                                            cand.instruction
+                                                        ),
+                                                        cand.score !== null && React.createElement('div', { className: 'text-xs text-gray-500 mb-3' },
+                                                            `评分: ${cand.score.toFixed(4)}`
+                                                        ),
+                                                        React.createElement('div', { className: 'flex gap-2' },
+                                                            React.createElement('button', {
+                                                                onClick: handleContinueOptimize,
+                                                                disabled: loading,
+                                                                className: 'px-4 py-2 bg-white border border-gray-300 text-gray-700 rounded-lg hover:bg-gray-50 disabled:bg-gray-100 disabled:text-gray-400 disabled:cursor-not-allowed transition-colors text-sm font-medium'
+                                                            }, '继续优化'),
+                                                            React.createElement('button', {
+                                                                onClick: () => handleCopyInstruction(cand.instruction),
+                                                                className: 'px-4 py-2 bg-white border border-gray-300 text-gray-700 rounded-lg hover:bg-gray-50 transition-colors text-sm font-medium flex items-center gap-1'
+                                                            },
+                                                                React.createElement('svg', { width: '16', height: '16', viewBox: '0 0 24 24', fill: 'none', stroke: 'currentColor', strokeWidth: '2' },
+                                                                    React.createElement('rect', { x: '9', y: '9', width: '13', height: '13', rx: '2', ry: '2' }),
+                                                                    React.createElement('path', { d: 'M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1' })
+                                                                ),
+                                                                '复制'
+                                                            ),
+                                                            React.createElement('button', {
+                                                                onClick: () => handleExecute(cand.instruction),
+                                                                disabled: loading,
+                                                                className: 'px-4 py-2 bg-gray-900 text-white rounded-lg hover:bg-gray-800 disabled:bg-gray-300 disabled:cursor-not-allowed transition-colors text-sm font-medium'
+                                                            }, '执行此指令')
+                                                        )
+                                                    )
+                                                )
+                                            )
+                                        )
+                                    )
+                                );
+                            } else if (msg.type === 'execution') {
+                                return React.createElement('div', { key: idx, className: 'flex justify-start' },
+                                    React.createElement('div', { className: 'max-w-2xl bg-gray-50 border border-gray-200 rounded-2xl p-5' },
+                                        React.createElement('div', { className: 'text-xs text-gray-600 mb-2 font-medium' },
+                                            '执行结果'
+                                        ),
+                                        React.createElement('div', { className: 'text-gray-800 whitespace-pre-wrap leading-relaxed' },
+                                            msg.response
+                                        )
+                                    )
+                                );
+                            }
+                        }),
+                        loading && React.createElement('div', { className: 'flex justify-start' },
+                            React.createElement('div', { className: 'bg-gray-100 rounded-2xl px-5 py-3 text-gray-600' },
+                                React.createElement('span', { className: 'loading-dots' }, '思考中')
+                            )
+                        ),
+                        React.createElement('div', { ref: chatEndRef })
+                    ),
+
+                    // Input Area
+                    React.createElement('div', { className: 'p-6 bg-white max-w-4xl mx-auto w-full' },
+                        React.createElement('div', { className: 'relative' },
+                            React.createElement('div', { className: 'bg-white border border-gray-300 rounded-3xl shadow-sm hover:shadow-md transition-shadow focus-within:shadow-md focus-within:border-gray-400' },
+                                // Textarea
+                                React.createElement('textarea', {
+                                    value: inputValue,
+                                    onChange: (e) => setInputValue(e.target.value),
+                                    onKeyPress: (e) => {
+                                        if (e.key === 'Enter' && !e.shiftKey) {
+                                            e.preventDefault();
+                                            handleSendMessage();
+                                        }
+                                    },
+                                    placeholder: currentRunId ? '输入消息...' : '在此输入提示词',
+                                    disabled: loading,
+                                    rows: 3,
+                                    className: 'w-full px-5 pt-4 pb-2 bg-transparent focus:outline-none disabled:bg-transparent text-gray-800 placeholder-gray-500 resize-none'
+                                }),
+                                // Toolbar
+                                React.createElement('div', { className: 'flex items-center justify-between px-4 pb-3 pt-1 border-t border-gray-100' },
+                                    // Left side - Model selector
+                                    React.createElement('div', { className: 'flex items-center gap-2' },
+                                        React.createElement('label', { className: 'text-xs text-gray-600' }, '模型:'),
+                                        React.createElement('select', {
+                                            value: selectedModel,
+                                            onChange: (e) => setSelectedModel(e.target.value),
+                                            className: 'text-sm px-2 py-1 border border-gray-300 rounded-lg bg-white text-gray-700 focus:outline-none focus:border-gray-400 cursor-pointer'
+                                        },
+                                            models.map(model =>
+                                                React.createElement('option', { key: model, value: model }, model)
+                                            )
+                                        )
+                                    ),
+                                    // Right side - Send button
+                                    React.createElement('button', {
+                                        onClick: handleSendMessage,
+                                        disabled: loading || !inputValue.trim(),
+                                        className: 'p-2.5 bg-gray-100 text-gray-700 rounded-full hover:bg-gray-200 disabled:bg-gray-50 disabled:text-gray-300 disabled:cursor-not-allowed transition-colors flex items-center justify-center'
+                                    },
+                                        React.createElement('svg', {
+                                            width: '20',
+                                            height: '20',
+                                            viewBox: '0 0 24 24',
+                                            fill: 'currentColor'
+                                        },
+                                            React.createElement('path', { d: 'M2.01 21L23 12 2.01 3 2 10l15 2-15 2z' })
+                                        )
+                                    )
+                                )
+                            ),
+                            !currentRunId && React.createElement('div', { className: 'text-xs text-gray-500 mt-3 px-4' },
+                                '输入任务描述后，AI 将为你生成优化的系统指令'
+                            )
+                        )
+                    )
+                )
+            );
+        }
+
+        // Render App
+        const root = ReactDOM.createRoot(document.getElementById('root'));
+        root.render(React.createElement(App));
+    </script>
+</body>
+</html>
+

test_opro_api.py

@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+"""
+Test script for TRUE OPRO API endpoints.
+
+This script tests the complete OPRO workflow:
+1. Create OPRO run
+2. Generate initial candidates
+3. Evaluate candidates
+4. Generate optimized candidates
+5. View results
+
+Usage:
+    python test_opro_api.py
+"""
+
+import requests
+import json
+import time
+
+BASE_URL = "http://127.0.0.1:8010"
+
+def print_section(title):
+    """Print a section header."""
+    print("\n" + "=" * 60)
+    print(f"  {title}")
+    print("=" * 60)
+
+def test_opro_workflow():
+    """Test the complete OPRO workflow."""
+    
+    print_section("1. Create OPRO Run")
+    
+    # Create a new OPRO run
+    create_req = {
+        "task_description": "将用户输入的中文翻译成英文，要求准确自然",
+        "test_cases": [
+            {"input": "你好", "expected_output": "Hello"},
+            {"input": "谢谢", "expected_output": "Thank you"},
+            {"input": "早上好", "expected_output": "Good morning"},
+            {"input": "晚安", "expected_output": "Good night"},
+            {"input": "再见", "expected_output": "Goodbye"}
+        ]
+    }
+    
+    response = requests.post(f"{BASE_URL}/opro/create", json=create_req)
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to create OPRO run: {result}")
+        return
+    
+    run_id = result["data"]["run_id"]
+    print(f"✅ Created OPRO run: {run_id}")
+    print(f"   Task: {result['data']['task_description']}")
+    print(f"   Test cases: {result['data']['num_test_cases']}")
+    
+    # ========================================================================
+    print_section("2. Generate Initial Candidates")
+    
+    iterate_req = {"run_id": run_id, "top_k": 5}
+    response = requests.post(f"{BASE_URL}/opro/iterate", json=iterate_req)
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to generate candidates: {result}")
+        return
+    
+    candidates = result["data"]["candidates"]
+    print(f"✅ Generated {len(candidates)} initial candidates:")
+    for i, candidate in enumerate(candidates, 1):
+        print(f"\n   [{i}] {candidate[:100]}...")
+    
+    # ========================================================================
+    print_section("3. Evaluate Candidates")
+    
+    scores = []
+    for i, candidate in enumerate(candidates, 1):
+        print(f"\n   Evaluating candidate {i}/{len(candidates)}...")
+        
+        eval_req = {
+            "run_id": run_id,
+            "instruction": candidate
+        }
+        
+        response = requests.post(f"{BASE_URL}/opro/evaluate", json=eval_req)
+        result = response.json()
+        
+        if result.get("success"):
+            score = result["data"]["score"]
+            scores.append(score)
+            is_best = "🏆" if result["data"]["is_new_best"] else ""
+            print(f"   ✅ Score: {score:.4f} {is_best}")
+        else:
+            print(f"   ❌ Evaluation failed: {result}")
+        
+        time.sleep(0.5)  # Small delay to avoid overwhelming the API
+    
+    print(f"\n   Average score: {sum(scores)/len(scores):.4f}")
+    print(f"   Best score: {max(scores):.4f}")
+    
+    # ========================================================================
+    print_section("4. Generate Optimized Candidates (Iteration 2)")
+    
+    print("   Generating candidates based on performance trajectory...")
+    
+    iterate_req = {"run_id": run_id, "top_k": 5}
+    response = requests.post(f"{BASE_URL}/opro/iterate", json=iterate_req)
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to generate optimized candidates: {result}")
+        return
+    
+    optimized_candidates = result["data"]["candidates"]
+    print(f"✅ Generated {len(optimized_candidates)} optimized candidates:")
+    for i, candidate in enumerate(optimized_candidates, 1):
+        print(f"\n   [{i}] {candidate[:100]}...")
+    
+    # ========================================================================
+    print_section("5. View Run Details")
+    
+    response = requests.get(f"{BASE_URL}/opro/run/{run_id}")
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to get run details: {result}")
+        return
+    
+    data = result["data"]
+    print(f"✅ OPRO Run Details:")
+    print(f"   Run ID: {data['run_id']}")
+    print(f"   Task: {data['task_description']}")
+    print(f"   Iteration: {data['iteration']}")
+    print(f"   Status: {data['status']}")
+    print(f"   Best Score: {data['best_score']:.4f}")
+    print(f"\n   Best Instruction:")
+    print(f"   {data['best_instruction'][:200]}...")
+    
+    print(f"\n   Top 5 Trajectory:")
+    for i, item in enumerate(data['trajectory'][:5], 1):
+        print(f"   [{i}] Score: {item['score']:.4f}")
+        print(f"       {item['instruction'][:80]}...")
+    
+    # ========================================================================
+    print_section("6. List All Runs")
+    
+    response = requests.get(f"{BASE_URL}/opro/runs")
+    result = response.json()
+    
+    if result.get("success"):
+        runs = result["data"]["runs"]
+        print(f"✅ Total OPRO runs: {result['data']['total']}")
+        for run in runs:
+            print(f"\n   Run: {run['run_id']}")
+            print(f"   Task: {run['task_description'][:50]}...")
+            print(f"   Iteration: {run['iteration']}, Best Score: {run['best_score']:.4f}")
+    
+    print_section("✅ OPRO Workflow Test Complete!")
+    print(f"\nRun ID: {run_id}")
+    print("You can view details at:")
+    print(f"  {BASE_URL}/opro/run/{run_id}")
+
+
+if __name__ == "__main__":
+    print("=" * 60)
+    print("  TRUE OPRO API Test")
+    print("=" * 60)
+    print(f"\nBase URL: {BASE_URL}")
+    print("\nMake sure the API server is running:")
+    print("  uvicorn _qwen_xinference_demo.api:app --host 127.0.0.1 --port 8010")
+    print("\nStarting test in 3 seconds...")
+    time.sleep(3)
+    
+    try:
+        test_opro_workflow()
+    except requests.exceptions.ConnectionError:
+        print("\n❌ ERROR: Could not connect to API server")
+        print("Please start the server first:")
+        print("  uvicorn _qwen_xinference_demo.api:app --host 127.0.0.1 --port 8010")
+    except Exception as e:
+        print(f"\n❌ ERROR: {e}")
+        import traceback
+        traceback.print_exc()
+