| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477 |
- """
- Step3: 基于匹配节点生成灵感点
- 基于 Step1 的 Top1 匹配结果,以匹配到的人设要素作为锚点,
- 让 Agent 分析可以产生哪些灵感点
- """
- import os
- import sys
- import json
- import asyncio
- from pathlib import Path
- from agents import Agent, Runner, trace
- from agents.tracing.create import custom_span
- from lib.my_trace import set_trace_smith as set_trace
- from lib.client import get_model
- from lib.data_loader import load_persona_data, load_inspiration_list, select_inspiration
- # 模型配置
- MODEL_NAME = "google/gemini-2.5-pro"
- # ========== System Prompt ==========
- GENERATE_INSPIRATIONS_PROMPT = """
- # 任务
- 你是一个内容创作者,现在要从一个锚点要素出发,模拟灵感产生的完整思维路径。
- ## 输入说明
- - **<人设体系></人设体系>**: 完整的人设系统,包含所有可用的节点
- - **<锚点要素></锚点要素>**: 作为起点的人设要素
- - **<要素定义></要素定义>**: 该要素的完整定义
- - **<要素上下文></要素上下文>**: 该要素的上下文信息
- ## 思维路径规则
- 每个灵感的产生都需要完整的思维路径:**[锚点] → [维度] → [节点/联想]**
- 1. **[锚点]**: 必须从给定的锚点要素出发
- 2. **[维度]**: 从锚点特征自然产生的思维方向(动态的,根据锚点特点决定)
- - 例如:跨节点组合、具体场景联想、情感延伸、时间维度、夸张变形等
- 3. **[节点/联想]**:
- - **[节点]**: 人设体系中的其他节点
- - **[联想]**: 自由联想的具体内容
- - 可以有多个,用 + 连接
- ## 示例
- 假设锚点是"宠物迷惑行为":
- ```
- 路径: [锚点]宠物迷惑行为 → [维度]跨节点组合 → [节点]职场疲惫
- 灵感点: 猫咪上班后的呆滞眼神
- 路径: [锚点]宠物迷惑行为 → [维度]具体场景联想 → [联想]盯着空气
- 灵感点: 猫咪对着空气打拳
- 路径: [锚点]宠物迷惑行为 → [维度]跨节点组合 → [节点]职场疲惫 + [节点]谐音梗
- 灵感点: 猫咪"喵"班摸鱼
- ```
- ## 输出格式(严格JSON)
- ```json
- {
- "灵感点列表": [
- {
- "路径": "[锚点]xxx → [维度]xxx → [节点/联想]xxx",
- "灵感点": "具体的灵感点描述"
- },
- {
- "路径": "[锚点]xxx → [维度]xxx → [联想]xxx",
- "灵感点": "具体的灵感点描述"
- }
- ]
- }
- ```
- **要求**:
- 1. 生成 8-15 个灵感点
- 2. 路径必须严格按照 `[锚点] → [维度] → [节点/联想]` 格式
- 3. 维度要多样化,探索不同的思考方向
- 4. 每个灵感点简洁明确(一句话)
- 5. 灵感点要具体、可执行、符合人设风格
- 6. 必须严格按照 JSON 格式输出
- """.strip()
- def create_agent(model_name: str, prompt: str, name: str) -> Agent:
- """创建 Agent
- Args:
- model_name: 模型名称
- prompt: System prompt
- name: Agent 名称
- Returns:
- Agent 实例
- """
- agent = Agent(
- name=name,
- instructions=prompt,
- model=get_model(model_name),
- tools=[],
- )
- return agent
- def parse_json_response(response_content: str, default_value: dict = None) -> dict:
- """解析 JSON 响应
- Args:
- response_content: Agent 返回的响应内容
- default_value: 解析失败时的默认返回值
- Returns:
- 解析后的字典
- """
- try:
- # 如果响应包含在 markdown 代码块中,提取 JSON 部分
- if "```json" in response_content:
- json_start = response_content.index("```json") + 7
- json_end = response_content.index("```", json_start)
- json_text = response_content[json_start:json_end].strip()
- elif "```" in response_content:
- json_start = response_content.index("```") + 3
- json_end = response_content.index("```", json_start)
- json_text = response_content[json_start:json_end].strip()
- else:
- json_text = response_content.strip()
- return json.loads(json_text)
- except Exception as e:
- print(f"解析响应失败: {e}")
- return default_value if default_value else {}
- def format_persona_system(persona_data: dict) -> str:
- """格式化完整人设系统为文本
- Args:
- persona_data: 人设数据
- Returns:
- 格式化的人设系统文本
- """
- lines = ["# 人设系统"]
- # 处理三个部分:灵感点列表、目的点、关键点列表
- for section_key, section_title in [
- ("灵感点列表", "【灵感点】灵感的来源和性质"),
- ("目的点", "【目的点】创作的目的和价值导向"),
- ("关键点列表", "【关键点】内容的核心主体和表达方式")
- ]:
- section_data = persona_data.get(section_key, [])
- if not section_data:
- continue
- lines.append(f"\n## {section_title}\n")
- for perspective in section_data:
- perspective_name = perspective.get("视角名称", "")
- lines.append(f"\n### 视角:{perspective_name}")
- for pattern in perspective.get("模式列表", []):
- pattern_name = pattern.get("分类名称", "")
- pattern_def = pattern.get("核心定义", "")
- lines.append(f"\n 【一级】{pattern_name}")
- if pattern_def:
- lines.append(f" 定义:{pattern_def}")
- # 二级细分
- for sub in pattern.get("二级细分", []):
- sub_name = sub.get("分类名称", "")
- sub_def = sub.get("分类定义", "")
- lines.append(f" 【二级】{sub_name}:{sub_def}")
- return "\n".join(lines)
- def find_element_definition(persona_data: dict, element_name: str) -> str:
- """从人设数据中查找要素的定义
- Args:
- persona_data: 人设数据
- element_name: 要素名称
- Returns:
- 要素定义文本,如果未找到则返回空字符串
- """
- # 在灵感点列表中查找
- for section_key in ["灵感点列表", "目的点", "关键点列表"]:
- section_data = persona_data.get(section_key, [])
- for perspective in section_data:
- for pattern in perspective.get("模式列表", []):
- # 检查一级分类
- if pattern.get("分类名称", "") == element_name:
- definition = pattern.get("核心定义", "")
- if definition:
- return definition
- # 检查二级分类
- for sub in pattern.get("二级细分", []):
- if sub.get("分类名称", "") == element_name:
- return sub.get("分类定义", "")
- return ""
- def find_step1_file(persona_dir: str, inspiration: str, model_name: str) -> str:
- """查找 step1 输出文件
- Args:
- persona_dir: 人设目录
- inspiration: 灵感点名称
- model_name: 模型名称
- Returns:
- step1 文件路径
- Raises:
- SystemExit: 找不到文件时退出
- """
- step1_dir = os.path.join(persona_dir, "how", "灵感点", inspiration)
- model_name_short = model_name.replace("google/", "").replace("/", "_")
- step1_file_pattern = f"*_step1_*_{model_name_short}.json"
- step1_files = list(Path(step1_dir).glob(step1_file_pattern))
- if not step1_files:
- print(f"❌ 找不到 step1 输出文件")
- print(f"查找路径: {step1_dir}/{step1_file_pattern}")
- sys.exit(1)
- return str(step1_files[0])
- async def generate_inspirations_with_paths(
- persona_system_text: str,
- matched_element: str,
- element_definition: str,
- element_context: str
- ) -> list:
- """生成带思维路径的灵感点列表
- Args:
- persona_system_text: 完整人设系统文本
- matched_element: 匹配要素
- element_definition: 要素定义
- element_context: 要素上下文
- Returns:
- 灵感点列表 [{"路径": "...", "灵感点": "..."}, ...]
- """
- task_description = f"""## 本次任务
- <人设体系>
- {persona_system_text}
- </人设体系>
- <锚点要素>
- {matched_element}
- </锚点要素>
- <要素定义>
- {element_definition if element_definition else '无'}
- </要素定义>
- <要素上下文>
- {element_context}
- </要素上下文>
- 请从锚点要素出发,模拟灵感产生的完整思维路径,生成多个灵感点,严格按照 JSON 格式输出。"""
- messages = [{
- "role": "user",
- "content": [{"type": "input_text", "text": task_description}]
- }]
- agent = create_agent(MODEL_NAME, GENERATE_INSPIRATIONS_PROMPT, "Inspiration Path Generator")
- result = await Runner.run(agent, input=messages)
- parsed = parse_json_response(result.final_output, {"灵感点列表": []})
- return parsed.get("灵感点列表", [])
- async def process_step3_generate_inspirations(
- step1_top1: dict,
- persona_data: dict,
- current_time: str = None,
- log_url: str = None
- ) -> dict:
- """执行灵感生成分析(核心业务逻辑 - 单层路径生成)
- Args:
- step1_top1: step1 的 top1 匹配结果
- persona_data: 完整的人设数据
- current_time: 当前时间戳
- log_url: trace URL
- Returns:
- 生成结果字典
- """
- # 从 step1 结果中提取信息
- business_info = step1_top1.get("业务信息", {})
- input_info = step1_top1.get("输入信息", {})
- matched_element = business_info.get("匹配要素", "")
- element_context = input_info.get("A_Context", "")
- # 格式化人设系统
- persona_system_text = format_persona_system(persona_data)
- # 查找要素定义
- element_definition = find_element_definition(persona_data, matched_element)
- print(f"\n{'=' * 80}")
- print(f"Step3: 基于锚点生成灵感路径")
- print(f"{'=' * 80}")
- print(f"锚点要素: {matched_element}")
- print(f"要素定义: {element_definition if element_definition else '(未找到定义)'}")
- print(f"模型: {MODEL_NAME}\n")
- # 生成灵感点(单层,包含完整思维路径)
- with custom_span(name="生成带路径的灵感点", data={"锚点要素": matched_element}):
- inspirations = await generate_inspirations_with_paths(
- persona_system_text, matched_element, element_definition, element_context
- )
- print(f"\n{'=' * 80}")
- print(f"完成!共生成 {len(inspirations)} 个灵感点")
- print(f"{'=' * 80}\n")
- # 预览前3个
- if inspirations:
- print("预览前3个灵感点:")
- for i, item in enumerate(inspirations[:3], 1):
- print(f" {i}. 路径: {item.get('路径', '')}")
- print(f" 灵感: {item.get('灵感点', '')}")
- print()
- # 构建输出
- return {
- "元数据": {
- "current_time": current_time,
- "log_url": log_url,
- "model": MODEL_NAME,
- "步骤": "Step3: 基于匹配节点生成灵感点"
- },
- "锚点信息": {
- "人设要素": matched_element,
- "要素定义": element_definition if element_definition else "无",
- "要素上下文": element_context
- },
- "step1_结果": step1_top1,
- "灵感点列表": inspirations
- }
- async def main(current_time: str, log_url: str, force: bool = False):
- """主函数
- Args:
- current_time: 当前时间戳
- log_url: 日志链接
- force: 是否强制重新执行(跳过已存在文件检查)
- """
- # 解析命令行参数
- persona_dir = sys.argv[1] if len(sys.argv) > 1 else "data/阿里多多酱/out/人设_1110"
- inspiration_arg = sys.argv[2] if len(sys.argv) > 2 else "0"
- # 第三个参数:force(如果从命令行调用且有该参数,则覆盖函数参数)
- if len(sys.argv) > 3 and sys.argv[3] == "force":
- force = True
- print(f"{'=' * 80}")
- print(f"Step3: 基于匹配节点生成灵感点")
- print(f"{'=' * 80}")
- print(f"人设目录: {persona_dir}")
- print(f"灵感参数: {inspiration_arg}")
- # 加载数据
- persona_data = load_persona_data(persona_dir)
- inspiration_list = load_inspiration_list(persona_dir)
- # 选择灵感
- try:
- inspiration_index = int(inspiration_arg)
- if 0 <= inspiration_index < len(inspiration_list):
- test_inspiration = inspiration_list[inspiration_index]
- print(f"使用灵感[{inspiration_index}]: {test_inspiration}")
- else:
- print(f"❌ 灵感索引超出范围: {inspiration_index}")
- sys.exit(1)
- except ValueError:
- if inspiration_arg in inspiration_list:
- test_inspiration = inspiration_arg
- print(f"使用灵感: {test_inspiration}")
- else:
- print(f"❌ 找不到灵感: {inspiration_arg}")
- sys.exit(1)
- # 查找并加载 step1 结果
- step1_file = find_step1_file(persona_dir, test_inspiration, MODEL_NAME)
- step1_filename = os.path.basename(step1_file)
- step1_basename = os.path.splitext(step1_filename)[0]
- print(f"Step1 输入文件: {step1_file}")
- # 构建输出文件路径
- output_dir = os.path.join(persona_dir, "how", "灵感点", test_inspiration)
- model_name_short = MODEL_NAME.replace("google/", "").replace("/", "_")
- scope_prefix = step1_basename.split("_")[0]
- result_index = 0
- output_filename = f"{scope_prefix}_step3_top{result_index + 1}_生成灵感_{model_name_short}.json"
- output_file = os.path.join(output_dir, output_filename)
- # 检查文件是否已存在
- if not force and os.path.exists(output_file):
- print(f"\n✓ 输出文件已存在,跳过执行: {output_file}")
- print(f"提示: 如需重新执行,请添加 'force' 参数\n")
- return
- with open(step1_file, 'r', encoding='utf-8') as f:
- step1_data = json.load(f)
- actual_inspiration = step1_data.get("灵感", "")
- step1_results = step1_data.get("匹配结果列表", [])
- if not step1_results:
- print("❌ step1 结果为空")
- sys.exit(1)
- print(f"灵感: {actual_inspiration}")
- # 默认处理 top1
- selected_result = step1_results[result_index]
- print(f"处理第 {result_index + 1} 个匹配结果(Top{result_index + 1})\n")
- # 执行核心业务逻辑
- output = await process_step3_generate_inspirations(
- step1_top1=selected_result,
- persona_data=persona_data,
- current_time=current_time,
- log_url=log_url
- )
- # 在元数据中添加 step1 匹配索引
- output["元数据"]["step1_匹配索引"] = result_index + 1
- # 保存结果
- os.makedirs(output_dir, exist_ok=True)
- with open(output_file, 'w', encoding='utf-8') as f:
- json.dump(output, f, ensure_ascii=False, indent=2)
- # 输出统计信息
- stats = output.get("统计", {})
- print(f"\n{'=' * 80}")
- print(f"统计信息:")
- print(f" 外层维度数: {stats.get('外层维度数', 0)}")
- print(f" 内层维度数: {stats.get('内层维度数', 0)}")
- print(f" 灵感点总数: {stats.get('灵感点总数', 0)}")
- print(f"{'=' * 80}")
- print(f"\n完成!结果已保存到: {output_file}")
- if log_url:
- print(f"Trace: {log_url}\n")
- if __name__ == "__main__":
- # 设置 trace
- current_time, log_url = set_trace()
- # 使用 trace 上下文包裹整个执行流程
- with trace("Step3: 生成灵感点"):
- asyncio.run(main(current_time, log_url))
|