Add GPU support and improve Docker deployment

- Add GPU deployment support with NVIDIA runtime
  - Update Dockerfile.allinone with GPU environment variables
  - Add comprehensive GPU_DEPLOYMENT.md guide

- Make port 11434 (Ollama) optional for security
  - Update DEPLOYMENT.md with CPU and GPU deployment options
  - Simplify default docker run commands
  - Update healthcheck to only check web application

- Add memory requirements documentation
  - Create MEMORY_REQUIREMENTS.md with model comparison
  - Add build-8b.sh script for lower memory usage
  - Document OOM troubleshooting steps

- Improve Docker build process
  - Add BUILD_TROUBLESHOOTING.md for common issues
  - Add DISTRIBUTION.md for image distribution methods
  - Update .gitignore to exclude large binary files
  - Improve docker-entrypoint.sh with better diagnostics

- Update .dockerignore to include ollama-linux-amd64.tgz
- Add backup file exclusions to .gitignore

.dockerignore

@@ -0,0 +1,28 @@
+__pycache__
+*.pyc
+*.pyo
+*.pyd
+.Python
+*.so
+*.egg
+*.egg-info
+dist
+build
+.git
+.gitignore
+.vscode
+.idea
+*.md
+!README.md
+# Include pre-downloaded Ollama binary for offline build
+!ollama-linux-amd64.tgz
+local_docs
+examples
+outputs
+.DS_Store
+*.log
+.env
+.venv
+venv
+env
+

.gitignore

@@ -149,6 +149,16 @@ outputs/
 *.log
 local_docs/
 
+# Docker build artifacts (DO NOT commit these - they are huge!)
+ollama-models/
+*.tar
+ollama-linux-amd64.tgz
+system-prompt-optimizer-*.tar
+*.tar.gz
+
+# Backup files from scripts
+*.bak
+
 # Node modules (if any frontend dependencies)
 node_modules/
 package-lock.json

DEPLOYMENT.md

@@ -0,0 +1,461 @@
+# Docker 部署指南
+
+本文档说明如何在无外网访问的服务器上部署系统提示词优化工具。
+
+## 部署方案
+
+本项目提供两种部署方案：
+
+### 方案 A: All-in-One 镜像（推荐，适用于无外网服务器）
+
+**优点**：
+- 包含所有依赖：应用代码 + Ollama + LLM 模型
+- 一个镜像文件，部署简单
+- 无需在目标服务器上安装任何额外软件（除了 Docker）
+
+**缺点**：
+- 镜像文件很大（10-20GB）
+- 传输时间较长
+
+### 方案 B: 分离部署（适用于已有 Ollama 的服务器）
+
+**优点**：
+- 镜像文件较小（~500MB）
+- 可以复用现有的 Ollama 服务
+
+**缺点**：
+- 需要在目标服务器上单独安装和配置 Ollama
+- 需要手动下载模型
+
+---
+
+## 方案 A: All-in-One 部署（推荐）
+
+### 前置要求
+
+#### 在开发机器上（有外网访问）
+
+1. **Docker** 已安装
+2. **Ollama** 已安装并运行
+3. **磁盘空间**：至少 30GB 可用空间
+4. 已下载所需的 Ollama 模型：
+   - `qwen3:14b` (主模型，~8GB)
+   - `qwen3-embedding:4b` (嵌入模型，~2GB)
+
+#### 在目标服务器上（无外网访问）
+
+1. **Docker** 已安装
+2. **磁盘空间**：至少 25GB 可用空间
+
+### 部署步骤
+
+#### 步骤 1: 下载所需的 Ollama 模型
+
+在开发机器上，确保已下载所需模型：
+
+```bash
+# 下载主模型（约 8GB）
+ollama pull qwen3:14b
+
+# 下载嵌入模型（约 2GB）
+ollama pull qwen3-embedding:4b
+
+# 验证模型已下载
+ollama list
+```
+
+#### 步骤 2: 导出 Ollama 模型
+
+```bash
+# 运行导出脚本
+./export-ollama-models.sh
+```
+
+这将创建 `ollama-models/` 目录，包含所有模型文件。
+
+#### 步骤 3: 构建 All-in-One Docker 镜像
+
+```bash
+# 运行构建脚本（推荐）
+./build-allinone.sh
+
+# 或手动构建
+docker build -f Dockerfile.allinone -t system-prompt-optimizer:allinone .
+```
+
+**注意**：构建过程可能需要 10-30 分钟，取决于机器性能。
+
+#### 步骤 4: 导出 Docker 镜像
+
+如果使用 `build-allinone.sh`，镜像已自动导出。否则手动导出：
+
+```bash
+# 导出镜像（约 10-20GB）
+docker save -o system-prompt-optimizer-allinone.tar system-prompt-optimizer:allinone
+
+# 验证文件大小
+ls -lh system-prompt-optimizer-allinone.tar
+```
+
+#### 步骤 5: 传输到目标服务器
+
+使用 scp、U盘或其他方式传输镜像文件：
+
+```bash
+# 使用 scp（如果网络可达）
+scp system-prompt-optimizer-allinone.tar user@server:/path/
+
+# 或使用 rsync（支持断点续传）
+rsync -avP --progress system-prompt-optimizer-allinone.tar user@server:/path/
+
+# 或使用 U盘/移动硬盘物理传输
+```
+
+#### 步骤 6: 在目标服务器上加载镜像
+
+```bash
+# 加载镜像（需要几分钟）
+docker load -i system-prompt-optimizer-allinone.tar
+
+# 如果遇到权限错误，使用 sudo
+# sudo docker load -i system-prompt-optimizer-allinone.tar
+
+# 验证镜像已加载
+docker images | grep system-prompt-optimizer
+```
+
+#### 步骤 7: 启动服务
+
+**CPU 模式（默认）:**
+
+```bash
+# 启动容器（推荐：仅暴露 Web 端口）
+docker run -d \
+  --name system-prompt-optimizer \
+  -p 8010:8010 \
+  --restart unless-stopped \
+  system-prompt-optimizer:allinone
+
+# 查看启动日志
+docker logs -f system-prompt-optimizer
+```
+
+**GPU 模式（推荐，如果有 NVIDIA GPU）:**
+
+```bash
+# 使用所有可用 GPU（推荐）
+docker run -d \
+  --name system-prompt-optimizer \
+  --gpus all \
+  -p 8010:8010 \
+  --restart unless-stopped \
+  system-prompt-optimizer:allinone
+
+# 或指定特定 GPU
+docker run -d \
+  --name system-prompt-optimizer \
+  --gpus '"device=0"' \
+  -p 8010:8010 \
+  --restart unless-stopped \
+  system-prompt-optimizer:allinone
+
+# 查看启动日志
+docker logs -f system-prompt-optimizer
+```
+
+**GPU 部署前提条件**:
+- 已安装 NVIDIA 驱动 (`nvidia-smi` 可用)
+- 已安装 NVIDIA Container Toolkit
+- GPU 显存 ≥ 10GB (14b 模型) 或 ≥ 6GB (8b 模型)
+
+**详细 GPU 部署指南**: 参见 [GPU_DEPLOYMENT.md](GPU_DEPLOYMENT.md)
+
+**重要**：
+- 首次启动需要等待 30-60 秒（CPU）或 10-20 秒（GPU），Ollama 服务需要初始化
+- GPU 模式下推理速度提升 5-10 倍
+- 端口 11434 (Ollama) 是可选的，仅在需要外部访问 Ollama 时暴露
+- 不暴露 11434 更安全，因为 Ollama API 没有身份验证
+
+#### 步骤 8: 验证部署
+
+```bash
+# 等待服务启动（约 30-60 秒）
+sleep 60
+
+# 健康检查
+curl http://localhost:8010/health
+
+# 应该返回：
+# {"status":"ok","version":"0.1.0"}
+
+# 检查 Ollama 服务
+curl http://localhost:11434/api/tags
+
+# 检查可用模型
+curl http://localhost:8010/models
+
+# 访问 Web 界面
+# 浏览器打开: http://<服务器IP>:8010/ui/opro.html
+```
+
+---
+
+## 方案 B: 分离部署
+
+### 前置要求
+
+#### 在目标服务器上
+
+1. **Docker** 已安装
+2. **Ollama** 服务已安装并运行
+3. 已拉取所需的 Ollama 模型：
+   - `qwen3:14b` (主模型)
+   - `qwen3-embedding:4b` (嵌入模型)
+
+### 部署步骤
+
+#### 步骤 1: 构建应用镜像
+
+```bash
+# 在开发机器上构建
+docker build -t system-prompt-optimizer:latest .
+
+# 导出镜像
+docker save -o system-prompt-optimizer.tar system-prompt-optimizer:latest
+```
+
+#### 步骤 2: 传输并加载
+
+```bash
+# 传输到目标服务器
+scp system-prompt-optimizer.tar user@server:/path/
+
+# 在目标服务器上加载
+docker load -i system-prompt-optimizer.tar
+```
+
+#### 步骤 3: 启动服务
+
+```bash
+# 使用 Docker Compose
+docker-compose up -d
+
+# 或使用 Docker 命令
+docker run -d \
+  --name system-prompt-optimizer \
+  -p 8010:8010 \
+  -e OLLAMA_HOST=http://host.docker.internal:11434 \
+  -v $(pwd)/outputs:/app/outputs \
+  --add-host host.docker.internal:host-gateway \
+  --restart unless-stopped \
+  system-prompt-optimizer:latest
+```
+
+## 配置说明
+
+### 环境变量
+
+在 `docker-compose.yml` 或 `docker run` 命令中可以配置以下环境变量：
+
+- `OLLAMA_HOST`: Ollama 服务地址（默认: `http://host.docker.internal:11434`）
+- `PYTHONUNBUFFERED`: Python 输出缓冲（默认: `1`）
+
+### 端口映射
+
+- **8010**: Web 界面和 API 端口（必需）
+- **11434**: Ollama API 端口（可选，仅用于调试或外部访问 Ollama）
+
+### 数据持久化
+
+- `./outputs`: 用户反馈日志存储目录（映射到容器内 `/app/outputs`）
+
+## 故障排查
+
+### 0. Docker 守护进程连接错误
+
+**问题**: 运行 `docker` 命令时提示 "Cannot connect to the Docker daemon"
+
+**症状**:
+```
+Cannot connect to the Docker daemon at unix:///var/run/docker.sock. Is the docker daemon running?
+```
+
+**解决方案**:
+
+**方法 1: 检查 Docker 服务状态**
+```bash
+# 检查 Docker 是否运行
+sudo systemctl status docker
+
+# 如果未运行，启动它
+sudo systemctl start docker
+
+# 设置开机自启
+sudo systemctl enable docker
+```
+
+**方法 2: 添加用户到 docker 组（推荐）**
+```bash
+# 将当前用户添加到 docker 组
+sudo usermod -aG docker $USER
+
+# 应用组变更（需要重新登录或使用 newgrp）
+newgrp docker
+
+# 或者直接注销并重新登录
+
+# 验证
+docker info
+```
+
+**方法 3: 修复 Docker socket 权限**
+```bash
+# 检查 socket 权限
+ls -l /var/run/docker.sock
+
+# 修复权限
+sudo chown root:docker /var/run/docker.sock
+sudo chmod 660 /var/run/docker.sock
+```
+
+**方法 4: 临时使用 sudo**
+```bash
+# 如果上述方法不可行，使用 sudo 运行 Docker 命令
+sudo docker load -i system-prompt-optimizer-allinone.tar
+sudo docker run -d --name system-prompt-optimizer ...
+```
+
+**验证修复**:
+```bash
+# 应该能正常显示 Docker 信息
+docker info
+
+# 应该能看到当前用户在 docker 组中
+groups | grep docker
+```
+
+---
+
+### 1. 无法连接 Ollama 服务
+
+**问题**: 容器内无法访问宿主机的 Ollama 服务
+
+**解决方案**:
+```bash
+# 确保使用了 --add-host 参数
+--add-host host.docker.internal:host-gateway
+
+# 或者直接使用宿主机 IP
+-e OLLAMA_HOST=http://192.168.1.100:11434
+```
+
+### 2. 模型不可用（All-in-One 部署）
+
+**问题**: 容器内模型未正确加载
+
+**解决方案**:
+```bash
+# 进入容器检查
+docker exec -it system-prompt-optimizer bash
+
+# 在容器内检查模型
+ollama list
+
+# 如果模型不存在，检查模型目录
+ls -la /root/.ollama/models/
+
+# 退出容器
+exit
+```
+
+如果模型确实丢失，可能需要重新构建镜像。
+
+### 3. 模型不可用（分离部署）
+
+**问题**: Ollama 模型未安装
+
+**解决方案**:
+```bash
+# 在宿主机上拉取模型
+ollama pull qwen3:14b
+ollama pull qwen3-embedding:4b
+
+# 验证模型已安装
+ollama list
+```
+
+### 4. 容器启动失败
+
+**问题**: 端口被占用或权限问题
+
+**解决方案**:
+```bash
+# 检查端口占用
+netstat -tulpn | grep 8010
+netstat -tulpn | grep 11434
+
+# 更换端口（All-in-One 需要两个端口）
+docker run -p 8011:8010 -p 11435:11434 ...
+
+# 查看容器日志
+docker logs system-prompt-optimizer
+```
+
+### 5. 性能问题
+
+**问题**: 生成速度慢
+
+**解决方案**:
+- 确保 Ollama 使用 GPU 加速
+- 使用更小的模型（如 `qwen3:4b`）
+- 调整 `config.py` 中的 `GENERATION_POOL_SIZE`
+
+## 更新部署
+
+```bash
+# 1. 在开发机器上重新构建镜像
+docker build -t system-prompt-optimizer:latest .
+
+# 2. 导出新镜像
+docker save -o system-prompt-optimizer-new.tar system-prompt-optimizer:latest
+
+# 3. 传输到服务器并加载
+docker load -i system-prompt-optimizer-new.tar
+
+# 4. 重启服务
+docker-compose down
+docker-compose up -d
+
+# 或使用 docker 命令
+docker stop system-prompt-optimizer
+docker rm system-prompt-optimizer
+docker run -d ...  # 使用相同的启动命令
+```
+
+## 安全建议
+
+1. **网络隔离**: 如果不需要外部访问，只绑定到 localhost
+   ```bash
+   -p 127.0.0.1:8010:8010
+   ```
+
+2. **防火墙**: 配置防火墙规则限制访问
+   ```bash
+   # 只允许特定 IP 访问
+   iptables -A INPUT -p tcp --dport 8010 -s 192.168.1.0/24 -j ACCEPT
+   iptables -A INPUT -p tcp --dport 8010 -j DROP
+   ```
+
+3. **日志管理**: 定期清理日志文件
+   ```bash
+   # 限制 Docker 日志大小
+   docker run --log-opt max-size=10m --log-opt max-file=3 ...
+   ```
+
+## 联系支持
+
+如有问题，请查看：
+- 应用日志: `docker logs system-prompt-optimizer`
+- Ollama 日志: `journalctl -u ollama -f`
+- API 文档: http://localhost:8010/docs
+

Dockerfile

@@ -0,0 +1,38 @@
+FROM python:3.10-slim
+
+# Set working directory
+WORKDIR /app
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy requirements file
+COPY requirements.txt .
+
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy application code
+COPY _qwen_xinference_demo/ ./_qwen_xinference_demo/
+COPY frontend/ ./frontend/
+COPY config.py .
+
+# Create outputs directory
+RUN mkdir -p outputs
+
+# Expose port
+EXPOSE 8010
+
+# Set environment variables
+ENV PYTHONUNBUFFERED=1
+ENV OLLAMA_HOST=http://host.docker.internal:11434
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost:8010/health || exit 1
+
+# Run the application
+CMD ["uvicorn", "_qwen_xinference_demo.api:app", "--host", "0.0.0.0", "--port", "8010"]
+

Dockerfile.allinone

@@ -0,0 +1,58 @@
+FROM --platform=linux/amd64 python:3.10-slim
+
+# Set working directory
+WORKDIR /app
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    curl \
+    ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Ollama manually for amd64
+# Copy pre-downloaded Ollama binary to avoid slow downloads during build
+# Using v0.13.1 (latest stable as of Dec 2024)
+COPY ollama-linux-amd64.tgz /tmp/ollama-linux-amd64.tgz
+RUN tar -C /usr -xzf /tmp/ollama-linux-amd64.tgz \
+    && rm /tmp/ollama-linux-amd64.tgz
+
+# Copy requirements file
+COPY requirements.txt .
+
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy application code
+COPY _qwen_xinference_demo/ ./_qwen_xinference_demo/
+COPY frontend/ ./frontend/
+COPY config.py .
+
+# Create necessary directories
+RUN mkdir -p outputs /root/.ollama
+
+# Copy pre-downloaded Ollama models
+# This includes qwen3:14b and qwen3-embedding:4b
+COPY ollama-models/ /root/.ollama/
+
+# Expose ports
+EXPOSE 8010 11434
+
+# Set environment variables
+ENV PYTHONUNBUFFERED=1
+ENV OLLAMA_HOST=http://localhost:11434
+# Enable GPU support for Ollama (will auto-detect NVIDIA GPU if available)
+ENV NVIDIA_VISIBLE_DEVICES=all
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+
+# Copy startup script
+COPY docker-entrypoint.sh /docker-entrypoint.sh
+RUN chmod +x /docker-entrypoint.sh
+
+# Health check
+# Only check the web application, not Ollama (internal service)
+HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
+    CMD curl -f http://localhost:8010/health || exit 1
+
+# Run the startup script
+ENTRYPOINT ["/docker-entrypoint.sh"]
+

QUICK_START.md

@@ -0,0 +1,117 @@
+# 快速开始指南
+
+## 离线部署（All-in-One 方案）
+
+### 在开发机器上（有外网）
+
+```bash
+# 1. 下载模型
+ollama pull qwen3:14b
+ollama pull qwen3-embedding:4b
+
+# 2. 导出模型
+./export-ollama-models.sh
+
+# 3. 构建并导出 Docker 镜像
+./build-allinone.sh
+
+# 4. 传输到目标服务器
+# 文件: system-prompt-optimizer-allinone.tar (约 10-20GB)
+scp system-prompt-optimizer-allinone.tar user@server:/path/
+```
+
+### 在目标服务器上（无外网）
+
+```bash
+# 1. 加载镜像
+docker load -i system-prompt-optimizer-allinone.tar
+
+# 2. 启动服务
+docker run -d \
+  --name system-prompt-optimizer \
+  -p 8010:8010 \
+  -p 11434:11434 \
+  -v $(pwd)/outputs:/app/outputs \
+  --restart unless-stopped \
+  system-prompt-optimizer:allinone
+
+# 3. 等待启动（约 60 秒）
+sleep 60
+
+# 4. 验证
+curl http://localhost:8010/health
+curl http://localhost:11434/api/tags
+
+# 5. 访问界面
+# http://<服务器IP>:8010/ui/opro.html
+```
+
+## 常用命令
+
+```bash
+# 查看日志
+docker logs -f system-prompt-optimizer
+
+# 重启服务
+docker restart system-prompt-optimizer
+
+# 停止服务
+docker stop system-prompt-optimizer
+
+# 删除容器
+docker rm -f system-prompt-optimizer
+
+# 进入容器
+docker exec -it system-prompt-optimizer bash
+
+# 检查模型
+docker exec -it system-prompt-optimizer ollama list
+```
+
+## 端口说明
+
+- **8010**: Web 界面和 API
+- **11434**: Ollama 服务（仅 All-in-One 方案需要暴露）
+
+## 文件说明
+
+- `system-prompt-optimizer-allinone.tar`: 完整镜像（10-20GB）
+- `outputs/`: 用户反馈日志目录
+
+## 故障排查
+
+### 服务无法启动
+
+```bash
+# 查看日志
+docker logs system-prompt-optimizer
+
+# 检查端口占用
+netstat -tulpn | grep 8010
+netstat -tulpn | grep 11434
+```
+
+### 模型不可用
+
+```bash
+# 进入容器检查
+docker exec -it system-prompt-optimizer ollama list
+
+# 应该看到:
+# qwen3:14b
+# qwen3-embedding:4b
+```
+
+### 性能慢
+
+- 确保服务器有足够的 RAM（建议 16GB+）
+- 如果有 GPU，使用支持 GPU 的 Docker 运行时
+- 调整 `config.py` 中的 `GENERATION_POOL_SIZE`
+
+## 更多信息
+
+详细文档请参考：
+- `DEPLOYMENT.md`: 完整部署指南
+- `README.md`: 项目说明
+- http://localhost:8010/docs: API 文档
+

README.md

@@ -1,58 +1,65 @@
-# OPRO Prompt Optimizer
+# System Prompt Generator
 
 ## 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。
+这是一个基于大语言模型的系统提示词（System Prompt）生成和迭代优化工具。通过简单的任务描述，自动生成高质量的系统指令，并支持基于用户选择的迭代改进。
 
 ### 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
+- **智能指令生成**：根据任务描述自动生成多个高质量的系统指令候选
-- **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
+- **迭代式改进**：基于用户选择的指令生成改进版本，避免被拒绝的方向
+- **角色定义格式**：所有生成的指令都以角色定义开头（如"你是一个..."），符合最佳实践
 - **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**：支持多个任务的并行管理和历史记录
-- **会话管理**：支持多个优化任务的并行管理和历史记录
+- **全面覆盖要求**：生成的指令全面覆盖任务的所有要求和细节，而非仅追求风格多样性
 
 ### 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
+- **实时生成反馈**：每轮生成 5 个候选指令，用户可选择继续优化或复制使用
 - **模型选择**：支持在界面中选择不同的 LLM 模型
 
-## 主要优化改进
+## 核心特性
 
-### 1. 真正的 OPRO 实现
+### 1. 简单直观的工作流程
 
-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：
+不同于复杂的 OPRO 算法（需要测试用例和自动评估），本工具采用简单直观的迭代改进方式：
 
-- **系统指令生成**：`generate_system_instruction_candidates()` - 生成多样化的系统指令候选
+- **初始生成**：输入任务描述 → 生成 5 个全面的系统指令候选
-- **性能评估**：`evaluate_system_instruction()` - 基于测试用例评估指令性能
+- **迭代改进**：选择喜欢的指令 → 生成基于该指令的改进版本，同时避免被拒绝的方向
-- **轨迹优化**：基于历史 (instruction, score) 轨迹生成更优指令
+- **无需评分**：不需要测试用例或性能评分，完全基于用户偏好进行改进
-- **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
+### 2. 高质量指令生成
 
-- **候选池大小优化**：从 20 个候选减少到 10 个，速度提升约 2 倍
+- **角色定义格式**：所有指令以"你是一个..."开头，符合系统提示词最佳实践
-- **智能聚类选择**：使用 AgglomerativeClustering 从候选池中选择最具多样性的 Top-K
+- **全面覆盖要求**：生成的指令全面覆盖任务的所有要求和细节
+- **清晰可执行**：指令清晰、具体、可执行，包含必要的行为规范和输出格式
+- **简体中文**：所有生成的指令使用简体中文
+
+### 3. 性能优化
+
+- **候选池大小优化**：生成 10 个候选，通过聚类选择 5 个最具多样性的
+- **智能聚类选择**：使用 AgglomerativeClustering 从候选池中选择最具代表性的指令
 - **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
+### 4. API 架构
 
-- **新增 OPRO 端点**：
+- **核心端点**：
-  - `POST /opro/create` - 创建 OPRO 优化任务
+  - `POST /opro/create` - 创建新任务
-  - `POST /opro/generate_and_evaluate` - 生成并自动评估候选
+  - `POST /opro/generate_and_evaluate` - 生成初始候选
-  - `POST /opro/execute` - 执行系统指令
+  - `POST /opro/refine` - 基于用户选择进行迭代改进
-  - `GET /opro/runs` - 获取所有优化任务
+  - `GET /opro/sessions` - 获取所有会话
-  - `GET /opro/run/{run_id}` - 获取特定任务详情
+  - `GET /opro/runs` - 获取所有任务
-- **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）
+- **会话管理**：支持多会话、多任务的并行管理
 - **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
+### 5. 前端界面
 
 - **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
+- **候选指令卡片**：每个候选显示编号和内容，提供"继续优化"和"复制"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
@@ -97,19 +104,18 @@ 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
+- **系统指令生成器**：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
+1. **创建新会话**：在界面点击"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"
+2. **输入任务描述**：例如"帮我写一个专业的营销文案生成助手"
-3. **查看候选指令**：系统生成 3-5 个优化的系统指令
+3. **查看候选指令**：系统生成 5 个全面的系统指令，每个都以角色定义开头
-4. **继续优化**：点击"继续优化"进行下一轮迭代
+4. **选择并改进**：点击喜欢的指令上的"继续优化"按钮，生成基于该指令的改进版本
-5. **执行指令**：点击"执行此指令"测试指令效果
+5. **复制使用**：点击"复制"按钮将指令复制到剪贴板，用于你的应用中
-6. **复制指令**：点击"复制"按钮将指令复制到剪贴板
 
 ## 配置说明
 
@@ -123,8 +129,8 @@ OLLAMA_HOST = "http://127.0.0.1:11434"
 DEFAULT_CHAT_MODEL = "qwen3:8b"
 DEFAULT_EMBED_MODEL = "qwen3-embedding:4b"
 
-# OPRO 优化参数
+# 生成参数
-GENERATION_POOL_SIZE = 10  # 生成候选池大小
+GENERATION_POOL_SIZE = 10  # 生成候选池大小（生成10个，聚类选择5个）
 TOP_K = 5                   # 返回给用户的候选数量
 CLUSTER_DISTANCE_THRESHOLD = 0.15  # 聚类距离阈值
 
@@ -157,11 +163,30 @@ XINFERENCE_EMBED_URL = "http://127.0.0.1:9997/models/bge-base-zh/embed"
 
 ## API 端点
 
-### OPRO 相关（推荐使用）
+### 会话管理
+
+- `POST /opro/session/create` - 创建新会话
+- `GET /opro/sessions` - 获取所有会话
+- `GET /opro/session/{session_id}` - 获取会话详情
+
+### 任务管理
+
+- `POST /opro/create` - 在会话中创建新任务
+  - 请求体：`{"session_id": "xxx", "task_description": "任务描述", "model_name": "qwen3:8b"}`
+  - 返回：`{"run_id": "xxx", "task_description": "...", "iteration": 0}`
+
+### 指令生成
+
+- `POST /opro/generate_and_evaluate` - 生成初始候选指令
+  - 请求体：`{"run_id": "xxx", "top_k": 5, "pool_size": 10}`
+  - 返回：`{"candidates": [{"instruction": "...", "score": null}, ...]}`
+
+- `POST /opro/refine` - 基于用户选择进行迭代改进
+  - 请求体：`{"run_id": "xxx", "selected_instruction": "用户选择的指令", "rejected_instructions": ["被拒绝的指令1", "被拒绝的指令2"]}`
+  - 返回：`{"candidates": [{"instruction": "...", "score": null}, ...], "iteration": 1}`
+
+### 任务查询
 
-- `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
 
@@ -181,6 +206,37 @@ XINFERENCE_EMBED_URL = "http://127.0.0.1:9997/models/bge-base-zh/embed"
 
 详细 API 文档请访问：http://127.0.0.1:8010/docs
 
+## 工作原理
+
+### 初始生成流程
+
+1. 用户输入任务描述（如"帮我写一个专业的营销文案生成助手"）
+2. 系统使用 LLM 生成 10 个候选指令
+3. 通过语义嵌入和聚类算法选择 5 个最具多样性的候选
+4. 所有候选都以角色定义开头，全面覆盖任务要求
+
+### 迭代改进流程
+
+1. 用户选择喜欢的指令（如候选 #3）
+2. 系统记录被拒绝的指令（候选 #1, #2, #4, #5）
+3. 向 LLM 发送改进请求："基于选中的指令生成改进版本，避免被拒绝指令的方向"
+4. 生成新的 10 个候选，聚类选择 5 个返回
+5. 用户可以继续迭代或复制使用
+
+### 与 OPRO 的区别
+
+**OPRO（原始算法）**：
+- 需要测试用例（如数学题的正确答案）
+- 自动评分（如准确率 0.73, 0.81）
+- 基于性能轨迹优化
+- 适用于有明确评估标准的任务
+
+**本工具（简单迭代改进）**：
+- 不需要测试用例
+- 不需要自动评分
+- 基于用户偏好改进
+- 适用于任意通用任务
+
 ## 常见问题
 
 ### 1. 无法连接 Ollama 服务
@@ -198,11 +254,17 @@ ollama serve
 
 ### 3. 生成速度慢
 
-- 调整 `GENERATION_POOL_SIZE` 减少候选数量
+- 调整 `GENERATION_POOL_SIZE` 减少候选数量（如改为 6，返回 3 个）
 - 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
+### 4. 生成的指令质量不高
+
+- 提供更详细的任务描述
+- 多次迭代改进，选择最好的继续优化
+- 尝试不同的模型
+
+### 5. 界面显示异常
 
 硬刷新浏览器缓存：
 - **Mac**: `Cmd + Shift + R`

_qwen_xinference_demo/api.py

@@ -14,6 +14,7 @@ from .opro.session_state import USER_FEEDBACK_LOG
 
 # True OPRO session management
 from .opro.session_state import (
+    create_opro_session, get_opro_session, list_opro_sessions,
     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
@@ -23,7 +24,8 @@ from .opro.session_state import (
 from .opro.user_prompt_optimizer import generate_candidates
 from .opro.user_prompt_optimizer import (
     generate_system_instruction_candidates,
-    evaluate_system_instruction
+    evaluate_system_instruction,
+    refine_instruction_candidates
 )
 
 from .opro.ollama_client import call_qwen
@@ -122,6 +124,7 @@ class CreateOPRORunReq(BaseModel):
     task_description: str
     test_cases: Optional[List[TestCase]] = None
     model_name: Optional[str] = None
+    session_id: Optional[str] = None  # Optional session to associate with
 
 
 class OPROIterateReq(BaseModel):
@@ -157,6 +160,15 @@ class OPROExecuteReq(BaseModel):
     model_name: Optional[str] = None
 
 
+class OPRORefineReq(BaseModel):
+    """Request to refine based on selected instruction (simple iterative refinement, NOT OPRO)."""
+    run_id: str
+    selected_instruction: str
+    rejected_instructions: List[str]
+    top_k: Optional[int] = None
+    pool_size: Optional[int] = None
+
+
 # ============================================================================
 # LEGACY ENDPOINTS (Query Rewriting - NOT true OPRO)
 # ============================================================================
@@ -360,12 +372,62 @@ def set_model(req: SetModelReq):
 # TRUE OPRO ENDPOINTS (System Instruction Optimization)
 # ============================================================================
 
+# Session Management
+
+@app.post("/opro/session/create", tags=["opro-true"])
+def opro_create_session(session_name: str = None):
+    """
+    Create a new OPRO session that can contain multiple runs.
+    """
+    session_id = create_opro_session(session_name=session_name)
+    session = get_opro_session(session_id)
+
+    return ok({
+        "session_id": session_id,
+        "session_name": session["session_name"],
+        "num_runs": len(session["run_ids"])
+    })
+
+
+@app.get("/opro/sessions", tags=["opro-true"])
+def opro_list_sessions():
+    """
+    List all OPRO sessions.
+    """
+    sessions = list_opro_sessions()
+    return ok({"sessions": sessions})
+
+
+@app.get("/opro/session/{session_id}", tags=["opro-true"])
+def opro_get_session(session_id: str):
+    """
+    Get detailed information about an OPRO session.
+    """
+    session = get_opro_session(session_id)
+    if not session:
+        raise AppException(404, "Session not found", "SESSION_NOT_FOUND")
+
+    # Get all runs in this session
+    runs = list_opro_runs(session_id=session_id)
+
+    return ok({
+        "session_id": session_id,
+        "session_name": session["session_name"],
+        "created_at": session["created_at"],
+        "num_runs": len(session["run_ids"]),
+        "runs": runs
+    })
+
+
+# Run Management
+
 @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.
+    Optionally can be associated with a session.
     """
     # Convert test cases from Pydantic models to tuples
     test_cases = None
@@ -375,7 +437,8 @@ def opro_create_run(req: CreateOPRORunReq):
     run_id = create_opro_run(
         task_description=req.task_description,
         test_cases=test_cases,
-        model_name=req.model_name
+        model_name=req.model_name,
+        session_id=req.session_id
     )
 
     run = get_opro_run(run_id)
@@ -385,7 +448,8 @@ def opro_create_run(req: CreateOPRORunReq):
         "task_description": run["task_description"],
         "num_test_cases": len(run["test_cases"]),
         "iteration": run["iteration"],
-        "status": run["status"]
+        "status": run["status"],
+        "session_id": run.get("session_id")
     })
 
 
@@ -433,15 +497,14 @@ def opro_evaluate(req: OPROEvaluateReq):
     Evaluate a system instruction on the test cases.
 
     This scores the instruction and updates the performance trajectory.
+    If no test cases are defined, uses a default score of 0.5 to indicate user selection.
     """
     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"]:
+    # Evaluate the instruction if test cases exist
-        raise AppException(400, "No test cases defined for this run", "NO_TEST_CASES")
+    if run["test_cases"] and len(run["test_cases"]) > 0:
-
-    # Evaluate the instruction
         try:
             score = evaluate_system_instruction(
                 system_instruction=req.instruction,
@@ -450,6 +513,10 @@ def opro_evaluate(req: OPROEvaluateReq):
             )
         except Exception as e:
             raise AppException(500, f"Evaluation failed: {e}", "EVALUATION_ERROR")
+    else:
+        # No test cases - use default score to indicate user selection
+        # This allows the trajectory to track which instructions the user preferred
+        score = 0.5
 
     # Add to trajectory
     add_opro_evaluation(req.run_id, req.instruction, score)
@@ -462,7 +529,8 @@ def opro_evaluate(req: OPROEvaluateReq):
         "instruction": req.instruction,
         "score": score,
         "best_score": run["best_score"],
-        "is_new_best": score == run["best_score"] and score > 0
+        "is_new_best": score == run["best_score"] and score > 0,
+        "has_test_cases": len(run["test_cases"]) > 0
     })
 
 
@@ -638,3 +706,44 @@ def opro_execute(req: OPROExecuteReq):
         })
     except Exception as e:
         raise AppException(500, f"Execution failed: {e}", "EXECUTION_ERROR")
+
+
+@app.post("/opro/refine", tags=["opro-true"])
+def opro_refine(req: OPRORefineReq):
+    """
+    Simple iterative refinement based on user selection (NOT OPRO).
+
+    This generates new candidates based on the selected instruction while avoiding rejected ones.
+    No scoring, no trajectory - just straightforward refinement based on user preference.
+    """
+    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
+
+    try:
+        candidates = refine_instruction_candidates(
+            task_description=run["task_description"],
+            selected_instruction=req.selected_instruction,
+            rejected_instructions=req.rejected_instructions,
+            top_k=top_k,
+            pool_size=pool_size,
+            model_name=run["model_name"]
+        )
+
+        # Update iteration counter
+        update_opro_iteration(req.run_id, candidates)
+
+        # Get updated run info
+        run = get_opro_run(req.run_id)
+
+        return ok({
+            "run_id": req.run_id,
+            "iteration": run["iteration"],
+            "candidates": [{"instruction": c, "score": None} for c in candidates],
+            "task_description": run["task_description"]
+        })
+    except Exception as e:
+        raise AppException(500, f"Refinement failed: {e}", "REFINEMENT_ERROR")

_qwen_xinference_demo/opro/prompt_utils.py

@@ -56,14 +56,15 @@ def generate_initial_system_instruction_candidates(task_description: str, pool_s
 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。
+请根据以上任务，生成 {pool_size} 条高质量、全面的"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
+1. 每条指令必须以角色定义开头（例如："你是一个..."、"你是..."等）
-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）
+2. 每条指令必须全面覆盖任务的所有要求和细节
-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
+3. 指令应清晰、具体、可执行，能够有效指导LLM完成任务
-4. 每条指令单独成行，不包含编号或额外说明
+4. 确保指令包含必要的行为规范、输出格式、注意事项等
-5. 所有生成的指令必须使用简体中文
+5. 每条指令单独成行，不包含编号或额外说明
+6. 所有生成的指令必须使用简体中文
 
 生成 {pool_size} 条指令：
 """
@@ -120,11 +121,68 @@ def generate_optimized_system_instruction(
 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
+1. 每条指令必须以角色定义开头（例如："你是一个..."、"你是..."等）
-2. 结合高分指令的优点，避免低分指令的缺陷
+2. 每条指令必须全面覆盖任务的所有要求和细节
-3. 探索新的优化方向和表达方式
+3. 结合高分指令的优点，避免低分指令的缺陷
-4. 每条指令单独成行，不包含编号或额外说明
+4. 指令应清晰、具体、可执行，能够有效指导LLM完成任务
-5. 所有生成的指令必须使用简体中文
+5. 每条指令单独成行，不包含编号或额外说明
+6. 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：
 """
+
+
+def refine_based_on_selection(
+    task_description: str,
+    selected_instruction: str,
+    rejected_instructions: List[str],
+    pool_size: int = None
+) -> str:
+    """
+    Simple refinement: Generate variations based on selected instruction while avoiding rejected ones.
+
+    This is NOT OPRO - it's straightforward iterative refinement based on user preference.
+    No scoring, no trajectory, just: "I like this one, give me more like it (but not like those)."
+
+    Args:
+        task_description: Description of the task
+        selected_instruction: The instruction the user selected
+        rejected_instructions: The instructions the user didn't select
+        pool_size: Number of new candidates to generate
+
+    Returns:
+        Prompt for generating refined candidates
+    """
+    import config
+    pool_size = pool_size or config.GENERATION_POOL_SIZE
+
+    rejected_text = ""
+    if rejected_instructions:
+        rejected_formatted = "\n".join(f"- {inst}" for inst in rejected_instructions)
+        rejected_text = f"""
+**用户未选择的指令（避免这些方向）：**
+{rejected_formatted}
+"""
+
+    return f"""
+你是一个"System Prompt 改进助手"。
+目标任务描述：
+【{task_description}】
+
+**用户选择的指令（基于此改进）：**
+{selected_instruction}
+{rejected_text}
+
+请基于用户选择的指令，生成 {pool_size} 条改进版本。
+
+要求：
+1. 每条指令必须以角色定义开头（例如："你是一个..."、"你是..."等）
+2. 保留用户选择指令的核心优点
+3. 每条指令必须全面覆盖任务的所有要求和细节
+4. 指令应清晰、具体、可执行，能够有效指导LLM完成任务
+5. 避免与未选择指令相似的方向
+6. 每条指令单独成行，不包含编号或额外说明
+7. 所有生成的指令必须使用简体中文
+
+生成 {pool_size} 条改进后的指令：
+"""

_qwen_xinference_demo/opro/session_state.py

@@ -66,10 +66,57 @@ def set_session_model(sid: str, model_name: str | None):
 # TRUE OPRO SESSION MANAGEMENT
 # ============================================================================
 
+# Session storage (contains multiple runs)
+OPRO_SESSIONS = {}
+
+def create_opro_session(session_name: str = None) -> str:
+    """
+    Create a new OPRO session that can contain multiple runs.
+
+    Args:
+        session_name: Optional name for the session
+
+    Returns:
+        session_id: Unique identifier for this session
+    """
+    session_id = uuid.uuid4().hex
+    OPRO_SESSIONS[session_id] = {
+        "session_name": session_name or "新会话",  # Will be updated with first task description
+        "created_at": uuid.uuid1().time,
+        "run_ids": [],  # List of run IDs in this session
+        "chat_history": []  # Cross-run chat history
+    }
+    return session_id
+
+
+def get_opro_session(session_id: str) -> Dict[str, Any]:
+    """Get OPRO session by ID."""
+    return OPRO_SESSIONS.get(session_id)
+
+
+def list_opro_sessions() -> List[Dict[str, Any]]:
+    """
+    List all OPRO sessions with summary information.
+
+    Returns:
+        List of session summaries
+    """
+    return [
+        {
+            "session_id": session_id,
+            "session_name": session["session_name"],
+            "num_runs": len(session["run_ids"]),
+            "created_at": session["created_at"]
+        }
+        for session_id, session in OPRO_SESSIONS.items()
+    ]
+
+
 def create_opro_run(
     task_description: str,
     test_cases: List[Tuple[str, str]] = None,
-    model_name: str = None
+    model_name: str = None,
+    session_id: str = None
 ) -> str:
     """
     Create a new OPRO optimization run.
@@ -78,6 +125,7 @@ def create_opro_run(
         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
+        session_id: Optional session ID to associate this run with
 
     Returns:
         run_id: Unique identifier for this OPRO run
@@ -87,6 +135,7 @@ def create_opro_run(
         "task_description": task_description,
         "test_cases": test_cases or [],
         "model_name": model_name,
+        "session_id": session_id,  # Link to parent session
         "iteration": 0,
         "trajectory": [],  # List of (instruction, score) tuples
         "best_instruction": None,
@@ -95,6 +144,14 @@ def create_opro_run(
         "created_at": uuid.uuid1().time,
         "status": "active"  # active, completed, failed
     }
+
+    # Add run to session if session_id provided
+    if session_id and session_id in OPRO_SESSIONS:
+        OPRO_SESSIONS[session_id]["run_ids"].append(run_id)
+        # Update session name with first task description if it's still default
+        if OPRO_SESSIONS[session_id]["session_name"] == "新会话" and len(OPRO_SESSIONS[session_id]["run_ids"]) == 1:
+            OPRO_SESSIONS[session_id]["session_name"] = task_description
+
     return run_id
 
 
@@ -206,13 +263,22 @@ def complete_opro_run(run_id: str):
         run["status"] = "completed"
 
 
-def list_opro_runs() -> List[Dict[str, Any]]:
+def list_opro_runs(session_id: str = None) -> List[Dict[str, Any]]:
     """
     List all OPRO runs with summary information.
 
+    Args:
+        session_id: Optional session ID to filter runs by session
+
     Returns:
         List of run summaries
     """
+    runs_to_list = OPRO_RUNS.items()
+
+    # Filter by session if provided
+    if session_id:
+        runs_to_list = [(rid, r) for rid, r in runs_to_list if r.get("session_id") == session_id]
+
     return [
         {
             "run_id": run_id,
@@ -220,7 +286,8 @@ def list_opro_runs() -> List[Dict[str, Any]]:
             "iteration": run["iteration"],
             "best_score": run["best_score"],
             "num_test_cases": len(run["test_cases"]),
-            "status": run["status"]
+            "status": run["status"],
+            "session_id": run.get("session_id")
         }
-        for run_id, run in OPRO_RUNS.items()
+        for run_id, run in runs_to_list
     ]

_qwen_xinference_demo/opro/user_prompt_optimizer.py

@@ -11,7 +11,8 @@ from .prompt_utils import (
     refine_instruction,
     refine_instruction_with_history,
     generate_initial_system_instruction_candidates,
-    generate_optimized_system_instruction
+    generate_optimized_system_instruction,
+    refine_based_on_selection
 )
 
 def parse_candidates(raw: str) -> list:
@@ -147,3 +148,46 @@ def evaluate_system_instruction(
             correct += 1
 
     return correct / total
+
+
+def refine_instruction_candidates(
+    task_description: str,
+    selected_instruction: str,
+    rejected_instructions: List[str],
+    top_k: int = config.TOP_K,
+    pool_size: int = None,
+    model_name: str = None
+) -> List[str]:
+    """
+    Simple refinement: Generate new candidates based on user's selection.
+
+    This is NOT OPRO - just straightforward iterative refinement.
+    User picks a favorite, we generate variations of it while avoiding rejected ones.
+
+    Args:
+        task_description: Description of the task
+        selected_instruction: The instruction the user selected
+        rejected_instructions: The instructions the user didn't select
+        top_k: Number of diverse candidates to return
+        pool_size: Number of candidates to generate before clustering
+        model_name: Optional model name to use
+
+    Returns:
+        List of refined instruction candidates
+    """
+    pool_size = pool_size or config.GENERATION_POOL_SIZE
+
+    # Generate the refinement prompt
+    meta_prompt = refine_based_on_selection(
+        task_description,
+        selected_instruction,
+        rejected_instructions,
+        pool_size
+    )
+
+    # Use LLM to generate refined candidates
+    raw = call_qwen(meta_prompt, temperature=0.9, max_tokens=1024, model_name=model_name)
+
+    # Parse and cluster
+    all_candidates = parse_candidates(raw)
+    return cluster_and_select(all_candidates, top_k=top_k)

build-8b.sh

@@ -0,0 +1,141 @@
+#!/bin/bash
+
+# Quick build script for qwen3:8b (lower memory usage)
+# Use this if your server has less than 12GB RAM
+
+set -e
+
+echo "=========================================="
+echo "Building with qwen3:8b (Lower Memory)"
+echo "=========================================="
+echo ""
+echo "Memory requirements:"
+echo "  - qwen3:8b: ~5GB RAM"
+echo "  - qwen3:14b: ~10GB RAM"
+echo ""
+
+# Check if 8b model is available
+if ! ollama list | grep -q "qwen3:8b"; then
+    echo "ERROR: qwen3:8b model not found!"
+    echo ""
+    echo "Please download it first:"
+    echo "  ollama pull qwen3:8b"
+    echo ""
+    exit 1
+fi
+
+# Clean up
+echo "Cleaning up previous builds..."
+rm -rf ollama-models/
+docker rmi system-prompt-optimizer:allinone 2>/dev/null || true
+
+# Export 8b model
+echo ""
+echo "Exporting qwen3:8b model..."
+mkdir -p ollama-models/models/{manifests/registry.ollama.ai/library,blobs}
+
+# Function to get blob hashes from manifest
+get_blobs_from_manifest() {
+    local manifest_file=$1
+    grep -o 'sha256:[a-f0-9]\{64\}' "$manifest_file" | sed 's/sha256://' | sort -u
+}
+
+# Function to copy model files
+copy_model() {
+    local model_name=$1
+    local model_tag=$2
+    local manifest_dir="$HOME/.ollama/models/manifests/registry.ollama.ai/library/$model_name"
+    
+    if [ ! -d "$manifest_dir" ]; then
+        echo "ERROR: Model manifest not found: $manifest_dir"
+        return 1
+    fi
+    
+    echo "  Copying $model_name:$model_tag manifest..."
+    mkdir -p "ollama-models/models/manifests/registry.ollama.ai/library/$model_name"
+    
+    if [ -f "$manifest_dir/$model_tag" ]; then
+        cp "$manifest_dir/$model_tag" "ollama-models/models/manifests/registry.ollama.ai/library/$model_name/"
+        
+        echo "  Finding blob files for $model_name:$model_tag..."
+        local blob_hashes=$(get_blobs_from_manifest "$manifest_dir/$model_tag")
+        local blob_count=0
+        
+        for blob_hash in $blob_hashes; do
+            local blob_file="$HOME/.ollama/models/blobs/sha256-$blob_hash"
+            if [ -f "$blob_file" ]; then
+                cp "$blob_file" "ollama-models/models/blobs/" 2>/dev/null
+                blob_count=$((blob_count + 1))
+            fi
+        done
+        
+        echo "  ✓ $model_name:$model_tag copied ($blob_count blobs)"
+    else
+        echo "ERROR: Manifest file not found: $manifest_dir/$model_tag"
+        return 1
+    fi
+}
+
+# Copy models
+copy_model "qwen3" "8b" || exit 1
+copy_model "qwen3-embedding" "4b" || exit 1
+
+echo ""
+echo "✓ Models exported successfully"
+echo ""
+
+# Update config.py to use 8b
+echo "Updating config.py to use qwen3:8b..."
+sed -i.bak 's/DEFAULT_CHAT_MODEL = "qwen3:14b"/DEFAULT_CHAT_MODEL = "qwen3:8b"/' config.py
+
+# Update docker-entrypoint.sh to check for 8b
+echo "Updating docker-entrypoint.sh to check for qwen3:8b..."
+sed -i.bak 's/qwen3:14b/qwen3:8b/g' docker-entrypoint.sh
+
+# Build image
+echo ""
+echo "Building Docker image..."
+docker build --platform linux/amd64 \
+    -f Dockerfile.allinone \
+    -t system-prompt-optimizer:allinone .
+
+if [ $? -ne 0 ]; then
+    echo ""
+    echo "Build failed!"
+    # Restore backups
+    mv config.py.bak config.py
+    mv docker-entrypoint.sh.bak docker-entrypoint.sh
+    exit 1
+fi
+
+# Export image
+echo ""
+echo "Exporting Docker image..."
+docker save -o system-prompt-optimizer-allinone.tar system-prompt-optimizer:allinone
+
+# Restore original files
+mv config.py.bak config.py
+mv docker-entrypoint.sh.bak docker-entrypoint.sh
+
+echo ""
+echo "=========================================="
+echo "Build Complete!"
+echo "=========================================="
+ls -lh system-prompt-optimizer-allinone.tar
+echo ""
+echo "This image uses qwen3:8b (~5GB RAM required)"
+echo ""
+echo "Transfer to server and run:"
+echo ""
+echo "  CPU mode:"
+echo "  docker load -i system-prompt-optimizer-allinone.tar"
+echo "  docker run -d -p 8010:8010 --restart unless-stopped system-prompt-optimizer:allinone"
+echo ""
+echo "  GPU mode (recommended):"
+echo "  docker load -i system-prompt-optimizer-allinone.tar"
+echo "  docker run -d --gpus all -p 8010:8010 --restart unless-stopped system-prompt-optimizer:allinone"
+echo ""
+echo "Note: GPU mode provides 5-10x faster inference."
+echo "      See GPU_DEPLOYMENT.md for GPU setup instructions."
+echo ""
+

build-allinone.sh

@@ -0,0 +1,133 @@
+#!/bin/bash
+
+# Build all-in-one Docker image with Ollama and models
+# This creates a complete offline-deployable image
+
+set -e
+
+IMAGE_NAME="system-prompt-optimizer"
+IMAGE_TAG="allinone"
+EXPORT_FILE="${IMAGE_NAME}-${IMAGE_TAG}.tar"
+
+echo "=========================================="
+echo "Building All-in-One Docker Image"
+echo "=========================================="
+echo ""
+echo "This will create a Docker image containing:"
+echo "  - Python application"
+echo "  - Ollama service (v0.13.1)"
+echo "  - qwen3:14b model"
+echo "  - qwen3-embedding:4b model"
+echo ""
+echo "Target platform: linux/amd64 (x86_64)"
+echo ""
+echo "WARNING: The final image will be 10-20GB in size!"
+echo ""
+echo "NOTE: If you're building on Apple Silicon (M1/M2/M3),"
+echo "      Docker will use emulation which may be slower."
+echo "      The image will still work on x86_64 servers."
+echo ""
+
+# Check if ollama-models directory exists
+if [ ! -d "ollama-models" ]; then
+    echo "ERROR: ollama-models directory not found!"
+    echo ""
+    echo "Please run ./export-ollama-models.sh first to export the models."
+    exit 1
+fi
+
+echo "✓ Found ollama-models directory"
+echo ""
+
+# Check if Ollama binary exists
+if [ ! -f "ollama-linux-amd64.tgz" ]; then
+    echo "ERROR: ollama-linux-amd64.tgz not found!"
+    echo ""
+    echo "Please download it first:"
+    echo "  curl -L -o ollama-linux-amd64.tgz https://github.com/ollama/ollama/releases/download/v0.13.1/ollama-linux-amd64.tgz"
+    echo ""
+    exit 1
+fi
+
+echo "✓ Found ollama-linux-amd64.tgz"
+echo ""
+
+# Check disk space
+AVAILABLE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+echo "Available disk space: $AVAILABLE_SPACE"
+echo "Required: ~20GB for build process"
+echo ""
+
+read -p "Continue with build? (y/n) " -n 1 -r
+echo
+if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+    echo "Build cancelled."
+    exit 1
+fi
+
+echo ""
+echo "=========================================="
+echo "Building Docker image..."
+echo "=========================================="
+echo "Platform: linux/amd64 (x86_64)"
+echo "This may take 20-40 minutes depending on your machine..."
+echo ""
+
+# Build for amd64 platform explicitly
+docker build --platform linux/amd64 -f Dockerfile.allinone -t ${IMAGE_NAME}:${IMAGE_TAG} .
+
+echo ""
+echo "=========================================="
+echo "Build complete!"
+echo "=========================================="
+docker images | grep ${IMAGE_NAME}
+
+echo ""
+echo "=========================================="
+echo "Exporting image to ${EXPORT_FILE}..."
+echo "=========================================="
+echo "This will take several minutes..."
+docker save -o ${EXPORT_FILE} ${IMAGE_NAME}:${IMAGE_TAG}
+
+echo ""
+echo "=========================================="
+echo "Export complete!"
+echo "=========================================="
+ls -lh ${EXPORT_FILE}
+
+echo ""
+echo "=========================================="
+echo "Deployment Instructions"
+echo "=========================================="
+echo ""
+echo "1. Transfer ${EXPORT_FILE} to target server:"
+echo "   scp ${EXPORT_FILE} user@server:/path/"
+echo ""
+echo "2. On target server, load the image:"
+echo "   docker load -i ${EXPORT_FILE}"
+echo ""
+echo "3. Run the container:"
+echo ""
+echo "   CPU mode:"
+echo "   docker run -d \\"
+echo "     --name system-prompt-optimizer \\"
+echo "     -p 8010:8010 \\"
+echo "     --restart unless-stopped \\"
+echo "     ${IMAGE_NAME}:${IMAGE_TAG}"
+echo ""
+echo "   GPU mode (recommended if NVIDIA GPU available):"
+echo "   docker run -d \\"
+echo "     --name system-prompt-optimizer \\"
+echo "     --gpus all \\"
+echo "     -p 8010:8010 \\"
+echo "     --restart unless-stopped \\"
+echo "     ${IMAGE_NAME}:${IMAGE_TAG}"
+echo ""
+echo "   Note: Port 11434 (Ollama) is optional and only needed for debugging."
+echo "   GPU mode provides 5-10x faster inference. See GPU_DEPLOYMENT.md for details."
+echo ""
+echo "4. Access the application:"
+echo "   http://<server-ip>:8010/ui/opro.html"
+echo ""
+echo "See DEPLOYMENT.md for more details."
+

build-and-export.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+
+# Build and export Docker image for offline deployment
+# Usage: ./build-and-export.sh
+
+set -e
+
+IMAGE_NAME="system-prompt-optimizer"
+IMAGE_TAG="latest"
+EXPORT_FILE="${IMAGE_NAME}.tar"
+
+echo "=========================================="
+echo "Building Docker image..."
+echo "=========================================="
+docker build -t ${IMAGE_NAME}:${IMAGE_TAG} .
+
+echo ""
+echo "=========================================="
+echo "Exporting Docker image to ${EXPORT_FILE}..."
+echo "=========================================="
+docker save -o ${EXPORT_FILE} ${IMAGE_NAME}:${IMAGE_TAG}
+
+echo ""
+echo "=========================================="
+echo "Export complete!"
+echo "=========================================="
+ls -lh ${EXPORT_FILE}
+
+echo ""
+echo "Next steps:"
+echo "1. Transfer ${EXPORT_FILE} to target server"
+echo "2. Transfer docker-compose.yml to target server (optional)"
+echo "3. On target server, run: docker load -i ${EXPORT_FILE}"
+echo "4. On target server, run: docker-compose up -d"
+echo ""
+echo "See DEPLOYMENT.md for detailed instructions."
+

config.py

@@ -7,7 +7,7 @@ APP_CONTACT = {"name": "OPRO Team", "url": "http://127.0.0.1:8010/ui/"}
 OLLAMA_HOST = "http://127.0.0.1:11434"
 OLLAMA_GENERATE_URL = f"{OLLAMA_HOST}/api/generate"
 OLLAMA_TAGS_URL = f"{OLLAMA_HOST}/api/tags"
-DEFAULT_CHAT_MODEL = "qwen3:8b"
+DEFAULT_CHAT_MODEL = "qwen3:14b"
 DEFAULT_EMBED_MODEL = "qwen3-embedding:4b"
 
 # Xinference

docker-compose.yml

@@ -0,0 +1,23 @@
+version: '3.8'
+
+services:
+  app:
+    build: .
+    container_name: system-prompt-optimizer
+    ports:
+      - "8010:8010"
+    environment:
+      - OLLAMA_HOST=http://host.docker.internal:11434
+      - PYTHONUNBUFFERED=1
+    volumes:
+      - ./outputs:/app/outputs
+    restart: unless-stopped
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8010/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 5s
+

docker-entrypoint.sh

@@ -0,0 +1,103 @@
+#!/bin/bash
+
+set -e
+
+echo "=========================================="
+echo "System Prompt Optimizer - Starting Up"
+echo "=========================================="
+echo ""
+
+# Check if Ollama binary exists
+if ! command -v ollama &> /dev/null; then
+    echo "ERROR: Ollama binary not found!"
+    echo "Expected location: /usr/bin/ollama or /usr/local/bin/ollama"
+    ls -la /usr/bin/ollama* 2>/dev/null || echo "No ollama in /usr/bin/"
+    ls -la /usr/local/bin/ollama* 2>/dev/null || echo "No ollama in /usr/local/bin/"
+    exit 1
+fi
+
+echo "✓ Ollama binary found: $(which ollama)"
+echo ""
+
+# Check if model files exist
+echo "Checking model files..."
+if [ ! -d "/root/.ollama/models" ]; then
+    echo "ERROR: /root/.ollama/models directory not found!"
+    exit 1
+fi
+
+MANIFEST_COUNT=$(find /root/.ollama/models/manifests -type f 2>/dev/null | wc -l)
+BLOB_COUNT=$(find /root/.ollama/models/blobs -type f 2>/dev/null | wc -l)
+
+echo "✓ Found $MANIFEST_COUNT manifest files"
+echo "✓ Found $BLOB_COUNT blob files"
+
+if [ "$BLOB_COUNT" -lt 10 ]; then
+    echo "WARNING: Very few blob files found. Models may not be complete."
+fi
+echo ""
+
+echo "Starting Ollama service..."
+ollama serve > /tmp/ollama.log 2>&1 &
+OLLAMA_PID=$!
+
+# Wait for Ollama to be ready
+echo "Waiting for Ollama to start..."
+OLLAMA_READY=false
+for i in {1..60}; do
+    if curl -s http://localhost:11434/api/tags > /dev/null 2>&1; then
+        echo "Ollama is ready!"
+        OLLAMA_READY=true
+        break
+    fi
+    echo "Waiting for Ollama... ($i/60)"
+    sleep 3
+done
+
+if [ "$OLLAMA_READY" = false ]; then
+    echo ""
+    echo "ERROR: Ollama failed to start within 3 minutes!"
+    echo ""
+    echo "Ollama logs:"
+    cat /tmp/ollama.log
+    echo ""
+    echo "Check full logs with: docker logs system-prompt-optimizer"
+    exit 1
+fi
+
+# Check if models exist, if not, show warning
+echo ""
+echo "Checking for models..."
+ollama list
+
+echo ""
+if ! ollama list | grep -q "qwen3:14b"; then
+    echo "ERROR: qwen3:14b model not found!"
+    echo "The application requires qwen3:14b to function properly."
+    echo ""
+    echo "Available models:"
+    ollama list
+    echo ""
+    exit 1
+fi
+
+if ! ollama list | grep -q "qwen3-embedding"; then
+    echo "WARNING: qwen3-embedding model not found!"
+    echo "The application requires qwen3-embedding:4b for embeddings."
+    echo "Continuing anyway, but embeddings may not work."
+fi
+
+echo ""
+echo "✓ All required models are available"
+echo ""
+echo "=========================================="
+echo "Starting FastAPI application..."
+echo "=========================================="
+echo "Application will be available at:"
+echo "  - Web UI: http://localhost:8010/ui/opro.html"
+echo "  - API Docs: http://localhost:8010/docs"
+echo "  - Ollama: http://localhost:11434"
+echo ""
+
+exec uvicorn _qwen_xinference_demo.api:app --host 0.0.0.0 --port 8010
+

export-ollama-models.sh

@@ -0,0 +1,168 @@
+#!/bin/bash
+
+# Export Ollama models for offline deployment
+# This script copies Ollama models from your local machine
+# so they can be bundled into the Docker image
+#
+# Required models:
+# - qwen3:14b (main chat model)
+# - qwen3-embedding:4b (embedding model)
+
+set -e
+
+MODELS_DIR="ollama-models"
+OLLAMA_MODELS_PATH="$HOME/.ollama"
+
+echo "=========================================="
+echo "Exporting Ollama models for offline deployment"
+echo "=========================================="
+
+# Check if Ollama is installed
+if ! command -v ollama &> /dev/null; then
+    echo "ERROR: Ollama is not installed or not in PATH"
+    exit 1
+fi
+
+# Check if required models are available
+echo ""
+echo "Checking for required models..."
+MISSING_MODELS=0
+
+if ! ollama list | grep -q "qwen3:14b"; then
+    echo "ERROR: qwen3:14b model not found!"
+    echo "Please run: ollama pull qwen3:14b"
+    MISSING_MODELS=1
+fi
+
+if ! ollama list | grep -q "qwen3-embedding:4b"; then
+    echo "ERROR: qwen3-embedding:4b model not found!"
+    echo "Please run: ollama pull qwen3-embedding:4b"
+    MISSING_MODELS=1
+fi
+
+if [ $MISSING_MODELS -eq 1 ]; then
+    echo ""
+    echo "Please download the required models first:"
+    echo "  ollama pull qwen3:14b"
+    echo "  ollama pull qwen3-embedding:4b"
+    exit 1
+fi
+
+echo "✓ All required models found"
+
+# Check if Ollama directory exists
+if [ ! -d "$OLLAMA_MODELS_PATH" ]; then
+    echo "ERROR: Ollama directory not found at $OLLAMA_MODELS_PATH"
+    exit 1
+fi
+
+# Create export directory structure
+echo ""
+echo "Creating export directory: $MODELS_DIR"
+rm -rf "$MODELS_DIR"
+mkdir -p "$MODELS_DIR/models/manifests/registry.ollama.ai/library"
+mkdir -p "$MODELS_DIR/models/blobs"
+
+echo ""
+echo "Copying only required models (qwen3:14b and qwen3-embedding:4b)..."
+echo "This may take several minutes (models are large)..."
+
+# Function to get blob hashes from manifest
+get_blobs_from_manifest() {
+    local manifest_file=$1
+    # Extract all sha256 hashes from the manifest JSON
+    grep -oE 'sha256:[a-f0-9]{64}' "$manifest_file" 2>/dev/null | sed 's/sha256://' | sort -u
+}
+
+# Function to copy model files
+copy_model() {
+    local model_name=$1
+    local model_tag=$2
+    local manifest_dir="$OLLAMA_MODELS_PATH/models/manifests/registry.ollama.ai/library/$model_name"
+
+    if [ ! -d "$manifest_dir" ]; then
+        echo "ERROR: Model manifest not found: $manifest_dir"
+        return 1
+    fi
+
+    echo "  Copying $model_name:$model_tag manifest..."
+    mkdir -p "$MODELS_DIR/models/manifests/registry.ollama.ai/library/$model_name"
+
+    # Copy the specific tag manifest
+    if [ -f "$manifest_dir/$model_tag" ]; then
+        cp "$manifest_dir/$model_tag" "$MODELS_DIR/models/manifests/registry.ollama.ai/library/$model_name/"
+
+        # Get all blob hashes referenced in this manifest
+        echo "  Finding blob files for $model_name:$model_tag..."
+        local blob_hashes=$(get_blobs_from_manifest "$manifest_dir/$model_tag")
+        local blob_count=0
+
+        for blob_hash in $blob_hashes; do
+            local blob_file="$OLLAMA_MODELS_PATH/models/blobs/sha256-$blob_hash"
+            if [ -f "$blob_file" ]; then
+                cp "$blob_file" "$MODELS_DIR/models/blobs/" 2>/dev/null
+                blob_count=$((blob_count + 1))
+            fi
+        done
+
+        echo "  ✓ $model_name:$model_tag copied ($blob_count blobs)"
+    else
+        echo "ERROR: Manifest file not found: $manifest_dir/$model_tag"
+        return 1
+    fi
+}
+
+# Copy required models with specific tags
+copy_model "qwen3" "14b" || exit 1
+copy_model "qwen3-embedding" "4b" || exit 1
+
+echo ""
+echo "=========================================="
+echo "Models exported successfully!"
+echo "=========================================="
+echo ""
+echo "Total size:"
+du -sh "$MODELS_DIR"
+
+echo ""
+echo "Models included:"
+if [ -d "$MODELS_DIR/models/manifests/registry.ollama.ai/library" ]; then
+    ls -lh "$MODELS_DIR/models/manifests/registry.ollama.ai/library/"
+fi
+
+echo ""
+echo "Blob files:"
+if [ -d "$MODELS_DIR/models/blobs" ]; then
+    echo "  Total blobs: $(ls -1 "$MODELS_DIR/models/blobs" | wc -l)"
+    du -sh "$MODELS_DIR/models/blobs"
+fi
+
+echo ""
+echo "=========================================="
+echo "Summary"
+echo "=========================================="
+echo "✓ Only qwen3:14b and qwen3-embedding:4b were exported"
+echo ""
+echo "Models in your Ollama that were NOT copied:"
+ollama list | grep -v "qwen3:14b" | grep -v "qwen3-embedding:4b" | tail -n +2 || echo "  (none)"
+echo ""
+echo "This keeps the Docker image size minimal!"
+
+echo ""
+echo "=========================================="
+echo "Next steps:"
+echo "=========================================="
+echo "1. Build the all-in-one Docker image:"
+echo "   ./build-allinone.sh"
+echo ""
+echo "2. Or manually:"
+echo "   docker build -f Dockerfile.allinone -t system-prompt-optimizer:allinone ."
+echo ""
+echo "3. Export the image:"
+echo "   docker save -o system-prompt-optimizer-allinone.tar system-prompt-optimizer:allinone"
+echo ""
+echo "4. Transfer to target server:"
+echo "   scp system-prompt-optimizer-allinone.tar user@server:/path/"
+echo ""
+echo "Note: The final Docker image will be very large (10-20GB) due to the models."
+

frontend/opro.html

@@ -6,7 +6,7 @@
     <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>
+    <title>系统提示词优化</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>
@@ -50,18 +50,22 @@
         // Main App Component
         function App() {
             const [sidebarOpen, setSidebarOpen] = useState(false);
-            const [runs, setRuns] = useState([]);
+            const [sessions, setSessions] = useState([]);
+            const [currentSessionId, setCurrentSessionId] = useState(null);
+            const [currentSessionRuns, setCurrentSessionRuns] = useState([]);
             const [currentRunId, setCurrentRunId] = useState(null);
             const [messages, setMessages] = useState([]);
+            const [sessionMessages, setSessionMessages] = useState({}); // Store messages per session
+            const [sessionLastRunId, setSessionLastRunId] = useState({}); // Store last run ID per session
             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
+            // Load sessions and models on mount
             useEffect(() => {
-                loadRuns();
+                loadSessions();
                 loadModels();
             }, []);
 
@@ -85,29 +89,78 @@
                 chatEndRef.current?.scrollIntoView({ behavior: 'smooth' });
             }, [messages]);
 
-            async function loadRuns() {
+            async function loadSessions() {
                 try {
-                    const res = await fetch(`${API_BASE}/opro/runs`);
+                    const res = await fetch(`${API_BASE}/opro/sessions`);
                     const data = await res.json();
                     if (data.success) {
-                        setRuns(data.data.runs || []);
+                        setSessions(data.data.sessions || []);
                     }
                 } catch (err) {
-                    console.error('Failed to load runs:', err);
+                    console.error('Failed to load sessions:', err);
+                }
+            }
+
+            async function loadSessionRuns(sessionId) {
+                try {
+                    const res = await fetch(`${API_BASE}/opro/session/${sessionId}`);
+                    const data = await res.json();
+                    if (data.success) {
+                        setCurrentSessionRuns(data.data.runs || []);
+                    }
+                } catch (err) {
+                    console.error('Failed to load session runs:', err);
+                }
+            }
+            
+            async function createNewSession() {
+                try {
+                    const res = await fetch(`${API_BASE}/opro/session/create`, {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' }
+                    });
+                    const data = await res.json();
+
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to create session');
+                    }
+
+                    const sessionId = data.data.session_id;
+                    setCurrentSessionId(sessionId);
+                    setCurrentSessionRuns([]);
+                    setCurrentRunId(null);
+                    setMessages([]);
+                    setSessionMessages(prev => ({ ...prev, [sessionId]: [] })); // Initialize empty messages for new session
+
+                    // Reload sessions list
+                    await loadSessions();
+
+                    return sessionId;
+                } catch (err) {
+                    alert('创建会话失败: ' + err.message);
+                    return null;
                 }
             }
 
             async function createNewRun(taskDescription) {
                 setLoading(true);
                 try {
-                    // Create run
+                    // Ensure we have a session
+                    let sessionId = currentSessionId;
+                    if (!sessionId) {
+                        sessionId = await createNewSession();
+                        if (!sessionId) return;
+                    }
+
+                    // Create run within session
                     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
+                            model_name: selectedModel || undefined,
+                            session_id: sessionId
                         })
                     });
                     const data = await res.json();
@@ -119,16 +172,33 @@
                     const runId = data.data.run_id;
                     setCurrentRunId(runId);
 
-                    // Add user message
+                    // Save this as the last run for this session
-                    setMessages([{ role: 'user', content: taskDescription }]);
+                    setSessionLastRunId(prev => ({
+                        ...prev,
+                        [sessionId]: runId
+                    }));
+
+                    // Add user message to existing messages (keep chat history)
+                    const newUserMessage = { role: 'user', content: taskDescription };
+                    setMessages(prev => {
+                        const updated = [...prev, newUserMessage];
+                        // Save to session messages
+                        setSessionMessages(prevSessions => ({
+                            ...prevSessions,
+                            [sessionId]: updated
+                        }));
+                        return updated;
+                    });
 
                     // Generate and evaluate candidates
                     await generateCandidates(runId);
 
-                    // Reload runs list
+                    // Reload sessions and session runs
-                    await loadRuns();
+                    await loadSessions();
+                    await loadSessionRuns(sessionId);
                 } catch (err) {
                     alert('创建任务失败: ' + err.message);
+                    console.error('Error creating run:', err);
                 } finally {
                     setLoading(false);
                 }
@@ -137,6 +207,7 @@
             async function generateCandidates(runId) {
                 setLoading(true);
                 try {
+                    console.log('Generating candidates for run:', runId);
                     const res = await fetch(`${API_BASE}/opro/generate_and_evaluate`, {
                         method: 'POST',
                         headers: { 'Content-Type': 'application/json' },
@@ -148,19 +219,33 @@
                     });
                     const data = await res.json();
 
+                    console.log('Generate candidates response:', data);
+
                     if (!data.success) {
                         throw new Error(data.error || 'Failed to generate candidates');
                     }
 
                     // Add assistant message with candidates
-                    setMessages(prev => [...prev, {
+                    const newAssistantMessage = {
                         role: 'assistant',
                         type: 'candidates',
                         candidates: data.data.candidates,
                         iteration: data.data.iteration
-                    }]);
+                    };
+                    setMessages(prev => {
+                        const updated = [...prev, newAssistantMessage];
+                        // Save to session messages
+                        if (currentSessionId) {
+                            setSessionMessages(prevSessions => ({
+                                ...prevSessions,
+                                [currentSessionId]: updated
+                            }));
+                        }
+                        return updated;
+                    });
                 } catch (err) {
                     alert('生成候选指令失败: ' + err.message);
+                    console.error('Error generating candidates:', err);
                 } finally {
                     setLoading(false);
                 }
@@ -185,12 +270,23 @@
                     }
 
                     // Add execution result
-                    setMessages(prev => [...prev, {
+                    const newExecutionMessage = {
                         role: 'assistant',
                         type: 'execution',
                         instruction: instruction,
                         response: data.data.response
-                    }]);
+                    };
+                    setMessages(prev => {
+                        const updated = [...prev, newExecutionMessage];
+                        // Save to session messages
+                        if (currentSessionId) {
+                            setSessionMessages(prevSessions => ({
+                                ...prevSessions,
+                                [currentSessionId]: updated
+                            }));
+                        }
+                        return updated;
+                    });
                 } catch (err) {
                     alert('执行失败: ' + err.message);
                 } finally {
@@ -204,25 +300,64 @@
 
                 setInputValue('');
 
-                if (!currentRunId) {
+                // Always create a new run with the message as task description
-                    // 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() {
+            async function handleContinueOptimize(selectedInstruction, allCandidates) {
-                if (!currentRunId || loading) return;
+                if (!currentRunId || loading || !selectedInstruction) return;
-                generateCandidates(currentRunId);
+
+                setLoading(true);
+                try {
+                    // Get rejected instructions (all except the selected one)
+                    const rejectedInstructions = allCandidates
+                        .map(c => c.instruction)
+                        .filter(inst => inst !== selectedInstruction);
+
+                    // Call the refinement endpoint
+                    const res = await fetch(`${API_BASE}/opro/refine`, {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({
+                            run_id: currentRunId,
+                            selected_instruction: selectedInstruction,
+                            rejected_instructions: rejectedInstructions
+                        })
+                    });
+                    const data = await res.json();
+
+                    if (!data.success) {
+                        throw new Error(data.error || 'Failed to refine instruction');
+                    }
+
+                    // Add refined candidates to messages
+                    const newMessage = {
+                        role: 'assistant',
+                        type: 'candidates',
+                        iteration: data.data.iteration,
+                        candidates: data.data.candidates
+                    };
+
+                    setMessages(prev => {
+                        const updated = [...prev, newMessage];
+                        // Save to session messages
+                        setSessionMessages(prevSessions => ({
+                            ...prevSessions,
+                            [currentSessionId]: updated
+                        }));
+                        return updated;
+                    });
+                } catch (err) {
+                    alert('优化失败: ' + err.message);
+                    console.error('Error refining instruction:', err);
+                } finally {
+                    setLoading(false);
+                }
             }
 
             function handleExecute(instruction) {
                 if (loading) return;
-                const userInput = prompt('请输入要处理的内容（可选）:');
+                executeInstruction(instruction, '');
-                executeInstruction(instruction, userInput);
             }
 
             function handleCopyInstruction(instruction) {
@@ -235,11 +370,33 @@
             }
 
             function handleNewTask() {
+                // Create new run within current session
                 setCurrentRunId(null);
                 setMessages([]);
                 setInputValue('');
             }
 
+            async function handleNewSession() {
+                // Create completely new session
+                const sessionId = await createNewSession();
+                if (sessionId) {
+                    setCurrentSessionId(sessionId);
+                    setCurrentSessionRuns([]);
+                    setCurrentRunId(null);
+                    setMessages([]);
+                    setInputValue('');
+                }
+            }
+
+            async function handleSelectSession(sessionId) {
+                setCurrentSessionId(sessionId);
+                // Restore the last run ID for this session
+                setCurrentRunId(sessionLastRunId[sessionId] || null);
+                // Load messages from session storage
+                setMessages(sessionMessages[sessionId] || []);
+                await loadSessionRuns(sessionId);
+            }
+
             async function loadRun(runId) {
                 setLoading(true);
                 try {
@@ -301,22 +458,22 @@
                     // 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)
+                            // New session button (expanded)
                             React.createElement('button', {
-                                onClick: handleNewTask,
+                                onClick: handleNewSession,
                                 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'),
+                            sessions.length > 0 && React.createElement('div', { className: 'text-xs text-gray-500 mb-2 px-2' }, '会话列表'),
-                            runs.map(run =>
+                            sessions.map(session =>
                                 React.createElement('div', {
-                                    key: run.run_id,
+                                    key: session.session_id,
-                                    onClick: () => loadRun(run.run_id),
+                                    onClick: () => handleSelectSession(session.session_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'
+                                        currentSessionId === session.session_id ? 'bg-gray-100' : 'hover:bg-gray-50'
                                     }`
                                 },
                                     React.createElement('svg', {
@@ -331,12 +488,12 @@
                                         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
+                                        session.session_name
                                     )
                                 )
                             )
                         ) : React.createElement('button', {
-                            onClick: handleNewTask,
+                            onClick: handleNewSession,
                             className: 'p-2 text-gray-600 hover:bg-gray-100 rounded-lg transition-colors flex items-center justify-center',
                             title: '新建会话'
                         },
@@ -352,7 +509,10 @@
                     // 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'
+                            '系统提示词优化'
+                        ),
+                        currentSessionId && React.createElement('div', { className: 'text-sm text-gray-500' },
+                            sessions.find(s => s.session_id === currentSessionId)?.session_name || '当前会话'
                         )
                     ),
 
@@ -391,7 +551,7 @@
                                                         ),
                                                         React.createElement('div', { className: 'flex gap-2' },
                                                             React.createElement('button', {
-                                                                onClick: handleContinueOptimize,
+                                                                onClick: () => handleContinueOptimize(cand.instruction, msg.candidates),
                                                                 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'),
@@ -404,12 +564,7 @@
                                                                     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')
                                                         )
                                                     )
                                                 )
@@ -452,7 +607,7 @@
                                             handleSendMessage();
                                         }
                                     },
-                                    placeholder: currentRunId ? '输入消息...' : '在此输入提示词',
+                                    placeholder: '输入任务描述，创建新的优化任务...',
                                     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'
@@ -489,8 +644,10 @@
                                     )
                                 )
                             ),
-                            !currentRunId && React.createElement('div', { className: 'text-xs text-gray-500 mt-3 px-4' },
+                            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'
+                                currentSessionId
+                                    ? '输入任务描述，AI 将为你生成优化的系统指令'
+                                    : '点击左侧"新建会话"开始，或直接输入任务描述自动创建会话'
                             )
                         )
                     )

requirements.txt

@@ -0,0 +1,7 @@
+fastapi==0.109.0
+uvicorn==0.27.0
+requests==2.31.0
+numpy==1.26.3
+scikit-learn==1.4.0
+pydantic==2.5.3
+

test_session_api.py

@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+"""
+Test script for OPRO session-based API
+"""
+
+import requests
+import json
+
+BASE_URL = "http://127.0.0.1:8010"
+
+def print_section(title):
+    print(f"\n{'='*60}")
+    print(f"  {title}")
+    print(f"{'='*60}\n")
+
+def test_session_workflow():
+    """Test the complete session-based workflow."""
+    
+    print_section("1. Create Session")
+    
+    # Create a new session
+    response = requests.post(f"{BASE_URL}/opro/session/create")
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to create session: {result}")
+        return
+    
+    session_id = result["data"]["session_id"]
+    print(f"✅ Session created: {session_id}")
+    print(f"   Session name: {result['data']['session_name']}")
+    
+    print_section("2. Create First Run in Session")
+    
+    # Create first run
+    create_req = {
+        "task_description": "将中文翻译成英文",
+        "test_cases": [
+            {"input": "你好", "expected_output": "Hello"},
+            {"input": "谢谢", "expected_output": "Thank you"}
+        ],
+        "session_id": session_id
+    }
+    
+    response = requests.post(f"{BASE_URL}/opro/create", json=create_req)
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to create run: {result}")
+        return
+    
+    run1_id = result["data"]["run_id"]
+    print(f"✅ Run 1 created: {run1_id}")
+    print(f"   Task: {result['data']['task_description']}")
+    
+    print_section("3. Create Second Run in Same Session")
+    
+    # Create second run in same session
+    create_req2 = {
+        "task_description": "将英文翻译成中文",
+        "test_cases": [
+            {"input": "Hello", "expected_output": "你好"},
+            {"input": "Thank you", "expected_output": "谢谢"}
+        ],
+        "session_id": session_id
+    }
+    
+    response = requests.post(f"{BASE_URL}/opro/create", json=create_req2)
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to create run 2: {result}")
+        return
+    
+    run2_id = result["data"]["run_id"]
+    print(f"✅ Run 2 created: {run2_id}")
+    print(f"   Task: {result['data']['task_description']}")
+    
+    print_section("4. Get Session Details")
+    
+    response = requests.get(f"{BASE_URL}/opro/session/{session_id}")
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to get session: {result}")
+        return
+    
+    print(f"✅ Session details:")
+    print(f"   Session ID: {result['data']['session_id']}")
+    print(f"   Session name: {result['data']['session_name']}")
+    print(f"   Number of runs: {result['data']['num_runs']}")
+    print(f"   Runs:")
+    for run in result['data']['runs']:
+        print(f"     - {run['run_id'][:8]}... : {run['task_description']}")
+    
+    print_section("5. List All Sessions")
+    
+    response = requests.get(f"{BASE_URL}/opro/sessions")
+    result = response.json()
+    
+    if not result.get("success"):
+        print(f"❌ Failed to list sessions: {result}")
+        return
+    
+    print(f"✅ Total sessions: {len(result['data']['sessions'])}")
+    for session in result['data']['sessions']:
+        print(f"   - {session['session_name']}: {session['num_runs']} runs")
+    
+    print_section("✅ All Tests Passed!")
+
+if __name__ == "__main__":
+    try:
+        # Check if server is running
+        response = requests.get(f"{BASE_URL}/health")
+        if response.status_code != 200:
+            print("❌ Server is not running. Please start it with:")
+            print("   uvicorn _qwen_xinference_demo.api:app --host 127.0.0.1 --port 8010")
+            exit(1)
+        
+        test_session_workflow()
+        
+    except requests.exceptions.ConnectionError:
+        print("❌ Cannot connect to server. Please start it with:")
+        print("   uvicorn _qwen_xinference_demo.api:app --host 127.0.0.1 --port 8010")
+        exit(1)
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        import traceback
+        traceback.print_exc()
+        exit(1)
+