from task_executor import TaskExecutor, TaskStatus from typing import Dict, Any import time import json import re class TextAnalysisExecutor(TaskExecutor): def __init__(self): super().__init__(llm_model="deepseek-chat") self.task_steps = [ { "id": "input_validation", "name": "Validate Input", "required_info": ["text"], "instruction": "Validate if the input text is not empty and contains valid Chinese characters. Check for proper UTF-8 encoding." }, { "id": "text_preprocessing", "name": "Preprocess Text", "required_info": ["text"], "instruction": "Clean and preprocess the text by: 1) Normalizing Chinese punctuation, 2) Removing unnecessary whitespace while preserving sentence structure, 3) Standardizing traditional/simplified characters if needed." }, { "id": "generate_summary", "name": "Generate Summary", "required_info": ["preprocessed_text"], "instruction": "Generate a concise summary in Chinese that captures the main points. Maintain the original language style and terminology." }, { "id": "extract_keywords", "name": "Extract Keywords", "required_info": ["preprocessed_text"], "instruction": "Extract the most important Chinese keywords and key phrases from the text. Include both technical terms and contextual phrases." }, { "id": "final_analysis", "name": "Final Analysis", "required_info": ["summary", "keywords"], "instruction": "Combine the summary and keywords into a comprehensive analysis report in Chinese. Structure the report with clear sections for summary, key points, and insights." } ] def validate_input(self, input_data: Dict[str, Any]) -> bool: """Validate specific input requirements for the text analysis task.""" if "text" not in input_data: return False text = input_data.get("text", "") return isinstance(text, str) and len(text.strip()) > 0 async def execute_step(self, step_id: str, step_data: Dict[str, Any]) -> bool: """Execute a specific step of the text analysis task.""" try: self.current_step = { "id": step_id, "name": step_data["name"], "status": TaskStatus.IN_PROGRESS.value, "progress": 0 } # Get step instruction and input instruction = step_data.get("instruction", "") step_input = {} # Prepare step-specific input if step_id == "input_validation": text = self.task_input.get("text", "") # Validate Chinese text encoding try: text.encode('utf-8').decode('utf-8') except UnicodeError: return {"error": "Invalid text encoding. Please ensure the text is properly encoded in UTF-8."} step_input = {"text": text} elif step_id == "text_preprocessing": text = self.task_input.get("text", "") # Basic Chinese text preprocessing # Normalize whitespace while preserving Chinese text structure text = re.sub(r'\s+', ' ', text).strip() # Normalize Chinese punctuation (simple example) text = text.replace(',', ',').replace('。', '.').replace(':', ':') step_input = {"text": text} elif step_id == "generate_summary": # Get the preprocessed text from previous step prev_result = next( (step["result"] for step in self.execution_path if step["step_id"] == "text_preprocessing"), {} ) step_input = {"text": prev_result.get("preprocessed_text", self.task_input.get("text", ""))} elif step_id == "extract_keywords": # Get the preprocessed text from previous step prev_result = next( (step["result"] for step in self.execution_path if step["step_id"] == "text_preprocessing"), {} ) step_input = {"text": prev_result.get("preprocessed_text", self.task_input.get("text", ""))} elif step_id == "final_analysis": # Get summary and keywords from previous steps summary_result = next( (step["result"] for step in self.execution_path if step["step_id"] == "generate_summary"), {} ) keywords_result = next( (step["result"] for step in self.execution_path if step["step_id"] == "extract_keywords"), {} ) step_input = { "summary": summary_result.get("summary", ""), "keywords": keywords_result.get("keywords", []) } # Execute step using LLM step_result = await self.llm_executor.execute_step( step_instruction=instruction, step_input=step_input ) if not step_result.get("success", False): raise Exception(f"Step failed: {step_result.get('error', 'Unknown error')}") # Update execution path self.execution_path.append({ "step_id": step_id, "result": step_result.get("output", {}) }) # Update step status self.current_step["status"] = TaskStatus.COMPLETED.value self.current_step["progress"] = 100 # Create checkpoint if needed if len(self.execution_path) % self.CHECKPOINT_INTERVAL == 0: self.create_checkpoint() return True except Exception as e: self.logger.error(f"Step {step_id} failed: {str(e)}") self.current_step["status"] = TaskStatus.FAILED.value return False def main(): # Set console encoding to UTF-8 import sys, io sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8') # Create the executor executor = TextAnalysisExecutor() # Sample Chinese text for analysis sample_text = """ 从 ChatGPT 到 Devin:AI 编程的四个发展阶段与范式转变 Koji:我们再聊一聊 AI 编程。编程领域今年取得了非常令人兴奋的进展。雨森一直有很强的框架归纳和总结能力。前不久你跟我分享过你提炼出来的 AI 编程发展四段论,要不要在播客里和大家分享一下? 雨森:这其实是和很多朋友一起探讨得出的结果,是大家智慧的结晶。AI 编程从 ChatGPT 出现到现在也就两年出头的时间,但已经经历了四个阶段。 第一个阶段是让 AI 直接写代码,典型代表是早期的 ChatGPT、Claude。我们给它一个需求,比如「帮我写个贪吃蛇」,它就给出一段代码。在这个过程中,它既不知道我为什么要写贪吃蛇,也不知道代码运行情况如何。可能要我去本地编译运行后发现报错,再把错误告诉它,它才能给出调试后的结果。这时的 AI 完全就像一个只能通过邮件交流的笔友,是简单的问答模式。 第二阶段是以 GitHub Copilot 为代表,AI 开始拥有上下文,它可以把整个组织的代码库作为 context。这样 AI 就获得了大量新的背景信息。但这时用户还是需要手动把代码贴到 IDE 里面进行调试。我觉得这是 2.0 阶段,就是我们让 AI 拥有了 codebase 作为上下文。 2024 年一个非常大的进步是以 Cursor 为代表的编程 Copilot 的出现。它的核心理念是预测用户未来要写什么代码。根据你的代码库以及刚才写的代码,它预测你接下来要写什么代码、创建什么文件、做什么操作。这里面对于生成代码的质量和数量,以及文件的创建和修改都有很大提升。后来 Windsurf 还加入了对命令行操作的自动化,这样 AI 就能很好地使用我的电脑。原来的 AI 是在一张纸上写代码,我把代码抄走运行;现在 AI 可以在我的电脑上创建文件、执行命令行操作,进入到「我为你写」的阶段。 当我们觉得这已经很令人兴奋时,Devin 的出现带来了几个重要突破:首先,它可以异步工作。Cursor、Windsurf 这些工具虽然一步操作做的事情比较多,但仍然需要持续的注意力,即「我说一步它做一步」。而 Devin 可以持续工作,把用户的注意力释放出来。这是因为它多了一个 Planner,可以规划任务。 其次,它可以通过虚拟机执行更多操作,做更多调试工作。比如你写个网站,它可以自己用虚拟机去访问这个网站,检查前端后端的业务逻辑是否正确,并且可以随时打断和调整。大家用 Cursor 或者 ChatGPT 都知道,你无法在它输出的中间做调整,必须等它输出完后才能修改。但 Devin 就像真人一样,你可以在它完成任务时给出新指令,它会把这个结合到已有的 Planner 里调整计划。这就从「为你写」进化到了「为你做」。 总结一下这四个阶段:第一阶段是让 AI 写代码,代表是 ChatGPT;第二阶段是 AI 开放代码库,代表是 GitHub Copilot;第三阶段是 AI 可以自动写代码并执行,代表是 Cursor 和 Windsurf;第四阶段是 AI 虚拟员工,Devin 开创了一个很好的先例。 """ # Prepare input task_input = { "text": sample_text } # Execute the task import asyncio result = asyncio.run(executor.execute(task_input)) # Print results with proper formatting print("\n 文本分析结果:") print("=" * 50) # Print each step's result with proper formatting for step in result.get("execution_path", []): step_id = step["step_id"] result_data = step["result"] print(f"\n {step_id.upper()}:") print("-" * 30) if step_id == "input_validation": print(" 输入验证完成") elif step_id == "text_preprocessing": print(" 文本预处理完成") if "preprocessed_text" in result_data: print("\n处理后的文本:") print(result_data["preprocessed_text"]) elif step_id == "generate_summary": print("\n 文本摘要:") if "summary" in result_data: print(result_data["summary"]) elif step_id == "extract_keywords": print("\n 关键词:") if "keywords" in result_data: keywords = result_data["keywords"] if isinstance(keywords, list): print("、".join(keywords)) else: print(keywords) elif step_id == "final_analysis": print("\n 最终分析:") if "analysis" in result_data: print(result_data["analysis"]) if __name__ == "__main__": main()