394 lines
20 KiB
Python
394 lines
20 KiB
Python
|
import os
|
|||
|
|
import json
|
|||
|
|
import random
|
|||
|
|
from typing import List, Dict, Tuple
|
|||
|
|
from openai import OpenAI
|
|||
|
|
from faker import Faker
|
|||
|
|
|
|||
|
|
class FullyDynamicGenerator:
|
|||
|
|
def __init__(self):
|
|||
|
|
self.llm = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
|
|||
|
|
self.faker = Faker('zh_CN')
|
|||
|
|
self.dynamic_memory = {}
|
|||
|
|
self.special_cases = [
|
|||
|
|
"方言沟通", "老年人口齿不清", "情绪激动打断对话",
|
|||
|
|
"背景噪音干扰", "信号断续"
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
def generate_dialog(self, category: str, subcategory: str) -> List[Dict]:
|
|||
|
|
"""全动态对话生成入口"""
|
|||
|
|
scene_knowledge = self.generate_scene_knowledge(category, subcategory)
|
|||
|
|
self.dynamic_memory[f"{category}_{subcategory}"] = scene_knowledge
|
|||
|
|
dialog = []
|
|||
|
|
dialog.extend(self._generate_complex_opening(category, subcategory))
|
|||
|
|
dialog.extend(self._generate_obstacle_base_phase(scene_knowledge, subcategory))
|
|||
|
|
dialog.extend(self._generate_verification_with_challenges(dialog))
|
|||
|
|
dialog.extend(self._generate_technical_extend_phase(scene_knowledge, subcategory))
|
|||
|
|
dialog.extend(self._generate_final_confirmation(scene_knowledge, subcategory))
|
|||
|
|
return self._format_output(dialog)
|
|||
|
|
|
|||
|
|
def _generate_complex_opening(self, category: str, subcategory: str) -> List[Tuple]:
|
|||
|
|
"""生成带复杂情形的开场对话"""
|
|||
|
|
phase = []
|
|||
|
|
special_case = random.choice(self.special_cases + [None]*3)
|
|||
|
|
|
|||
|
|
citizen_traits = {
|
|||
|
|
"方言": random.choice(["带浓重口音", "夹杂方言词汇", "语法不规范"]),
|
|||
|
|
"老年人": random.choice(["说话缓慢", "重复语句", "耳背听不清"]),
|
|||
|
|
"情绪化": random.choice(["不断打断", "提高音量", "带哭腔"])
|
|||
|
|
}
|
|||
|
|
opening_prompt = f"""生成市民反映{subcategory}问题的电话开场白，要求：
|
|||
|
|
1. 必须包含"您好"等礼貌用语
|
|||
|
|
2. 体现真实通话特征：{citizen_traits.get(special_case, "正常沟通")}
|
|||
|
|
3. 包含具体问题细节"""
|
|||
|
|
opening = self._safe_llm_call(
|
|||
|
|
prompt=opening_prompt,
|
|||
|
|
system="你擅长模拟各类人群的真实对话",
|
|||
|
|
response_format={"type": "json_object"}
|
|||
|
|
)
|
|||
|
|
try:
|
|||
|
|
opening_data = json.loads(opening)
|
|||
|
|
opening_text = opening_data.get("text", f"您好，我要反映{subcategory}问题")
|
|||
|
|
if special_case == "方言沟通":
|
|||
|
|
opening_text = self._add_dialect_features(opening_text)
|
|||
|
|
except:
|
|||
|
|
opening_text = f"您好，我想投诉{subcategory}问题"
|
|||
|
|
phase.append(("市民", "open_call", opening_text))
|
|||
|
|
response_prompt = f"""根据市民来电特征：{special_case if special_case else '正常'}，生成专业应答：
|
|||
|
|
1. 包含工号和服务承诺
|
|||
|
|
2. 适应沟通特征：{citizen_traits.get(special_case, '标准服务')}"""
|
|||
|
|
response = self._safe_llm_call(
|
|||
|
|
prompt=response_prompt,
|
|||
|
|
system="你是适应力强的专业客服",
|
|||
|
|
response_format={"type": "json_object"}
|
|||
|
|
)
|
|||
|
|
try:
|
|||
|
|
response_data = json.loads(response)
|
|||
|
|
response_text = response_data.get("text", f"感谢来电，工号{random.randint(1000,1999)}为您服务")
|
|||
|
|
if special_case == "老年人口齿不清":
|
|||
|
|
response_text += "（放慢语速）请您慢慢说"
|
|||
|
|
except:
|
|||
|
|
response_text = "您好，政务热线为您服务"
|
|||
|
|
phase.append(("客服", "agent_response", response_text))
|
|||
|
|
if special_case in ["方言沟通", "老年人口齿不清", "信号断续"]:
|
|||
|
|
phase.append(("客服", "double_check", f"抱歉，刚才没有听清楚，您是说{subcategory}问题对吗？"))
|
|||
|
|
phase.append(("市民", "clarify", random.choice([
|
|||
|
|
"对，就是这个问题",
|
|||
|
|
f"不是，是{random.choice(['更严重','其他'])}的问题",
|
|||
|
|
"（声音断断续续）喂...听得到吗？"
|
|||
|
|
])))
|
|||
|
|
return phase
|
|||
|
|
|
|||
|
|
def _generate_obstacle_base_phase(self, knowledge: Dict, scene: str) -> List[Tuple]:
|
|||
|
|
"""生成带沟通障碍的基础信息采集"""
|
|||
|
|
phase = []
|
|||
|
|
required_fields = ["时间", "地点", "事件描述", "联系方式", "姓氏"]
|
|||
|
|
for field in required_fields:
|
|||
|
|
if random.random() < 0.1:
|
|||
|
|
unclear_question = self._safe_llm_call(
|
|||
|
|
prompt=f"生成有歧义的{field}询问话术",
|
|||
|
|
system="故意制造1-2处不明确表述"
|
|||
|
|
) or f"那个...关于{field}的情况能不能说下？"
|
|||
|
|
phase.append(("客服", "unclear_question", unclear_question))
|
|||
|
|
phase.append(("市民", "confused", "您问的是什么？我没听明白"))
|
|||
|
|
question = self._safe_llm_call(
|
|||
|
|
prompt=f"重新生成清晰的{field}询问话术",
|
|||
|
|
system="使用最简明的表达"
|
|||
|
|
) or f"请提供{field}的具体信息"
|
|||
|
|
phase.append(("客服", "retry_question", question))
|
|||
|
|
else:
|
|||
|
|
question = self._safe_llm_call(
|
|||
|
|
prompt=f"生成政务热线询问{field}的标准话术，场景：{scene}",
|
|||
|
|
system="要求：1.使用敬语 2.明确信息要求"
|
|||
|
|
) or f"请问{scene}的{field}是？"
|
|||
|
|
phase.append(("客服", "info_request", question))
|
|||
|
|
answer, needs_clarify = self._generate_complex_answer(scene, field)
|
|||
|
|
phase.append(("市民", "info_response", answer))
|
|||
|
|
if needs_clarify:
|
|||
|
|
clarify_question = self._safe_llm_call(
|
|||
|
|
prompt=f"根据模糊回答'{answer}'生成澄清{field}的追问",
|
|||
|
|
system="要求：1.指出不明确处 2.提供填写范例"
|
|||
|
|
) or f"您提供的{field}不够具体，请补充（例：{self._get_field_example(field)}）"
|
|||
|
|
phase.append(("客服", "clarify_request", clarify_question))
|
|||
|
|
if random.random() < 0.1:
|
|||
|
|
phase.append(("市民", "refuse", random.choice([
|
|||
|
|
"这么麻烦不说了！",
|
|||
|
|
"你们政府办事就是繁琐",
|
|||
|
|
f"{field}有什么好问的！"
|
|||
|
|
])))
|
|||
|
|
phase.append(("客服", "calm_down", random.choice([
|
|||
|
|
"理解您的心情，但详细信息能帮助我们更快解决问题",
|
|||
|
|
"抱歉给您带来不便，这是必要流程"
|
|||
|
|
])))
|
|||
|
|
phase.append(("市民", "clarified_response", f"哦，应该是{self._get_field_example(field)}"))
|
|||
|
|
return phase
|
|||
|
|
|
|||
|
|
def _generate_complex_answer(self, scene: str, field: str) -> Tuple[str, bool]:
|
|||
|
|
"""生成带复杂特征的市民回答"""
|
|||
|
|
if random.random() < 0.15:
|
|||
|
|
special_answers = {
|
|||
|
|
"时间": [
|
|||
|
|
("就...就那个...前几天", True),
|
|||
|
|
("（背景嘈杂）喂？时间啊...上周？", True),
|
|||
|
|
("我不记得了！你们自己查！", False)
|
|||
|
|
],
|
|||
|
|
"地点": [
|
|||
|
|
("俺们村东头那个...那个啥来着", True),
|
|||
|
|
("（信号不好）在...哗哗...超市附近", True),
|
|||
|
|
("这么简单的问题都处理不了？", False)
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
return random.choice(special_answers.get(field, [("这个我说不好", True)]))
|
|||
|
|
answers = {
|
|||
|
|
"时间": [
|
|||
|
|
(f"{random.choice(['今天','昨天'])}{random.randint(1,12)}点左右", False),
|
|||
|
|
(f"持续{random.randint(2,24)}小时了", False)
|
|||
|
|
],
|
|||
|
|
"地点": [
|
|||
|
|
(f"{self.faker.building_number()}号{random.choice(['东侧','南门'])}", False),
|
|||
|
|
(f"{self.faker.street_name()}附近", True)
|
|||
|
|
],
|
|||
|
|
"联系方式": [
|
|||
|
|
(f"{self.faker.phone_number()[:3]}****", True),
|
|||
|
|
(f"固话：{self.faker.phone_number()[:4]}-{self.faker.phone_number()[-4:]}", False)
|
|||
|
|
],
|
|||
|
|
"姓氏": [
|
|||
|
|
(f"免贵姓{self.faker.last_name()}", False),
|
|||
|
|
("叫我老李就行", True)
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
return random.choice(answers.get(field, [("具体情况是这样的...", False)]))
|
|||
|
|
|
|||
|
|
def _generate_verification_with_challenges(self, previous_dialog: List[Tuple]) -> List[Tuple]:
|
|||
|
|
"""生成带挑战的信息确认环节"""
|
|||
|
|
phase = []
|
|||
|
|
collected_info = {}
|
|||
|
|
for turn in previous_dialog:
|
|||
|
|
if turn[1] in ["info_response", "clarified_response"]:
|
|||
|
|
for field in ["时间", "地点", "姓氏"]:
|
|||
|
|
if field in turn[2]:
|
|||
|
|
collected_info[field] = turn[2]
|
|||
|
|
if random.random() < 0.1:
|
|||
|
|
collected_info[field] = self._get_wrong_info(field)
|
|||
|
|
if collected_info:
|
|||
|
|
if random.random() < 0.05:
|
|||
|
|
wrong_field = random.choice(list(collected_info.keys()))
|
|||
|
|
correct_value = collected_info[wrong_field]
|
|||
|
|
collected_info[wrong_field] = self._get_wrong_info(wrong_field)
|
|||
|
|
verification_text = self._safe_llm_call(
|
|||
|
|
prompt="根据以下信息生成确认话术：" + json.dumps(collected_info, ensure_ascii=False),
|
|||
|
|
system="要求：1.逐项确认 2.允许修正"
|
|||
|
|
) or f"我确认下：时间：{collected_info.get('时间','')}，地点：{collected_info.get('地点','')}..."
|
|||
|
|
phase.append(("客服", "info_verification", verification_text))
|
|||
|
|
if random.random() < 0.3:
|
|||
|
|
correction_field = random.choice(list(collected_info.keys()))
|
|||
|
|
phase.append(("市民", "correction",
|
|||
|
|
f"{correction_field}不对！应该是{self._get_field_example(correction_field)}"))
|
|||
|
|
if random.random() < 0.1:
|
|||
|
|
phase.append(("市民", "angry", "你们连基本信息都记错！"))
|
|||
|
|
phase.append(("客服", "apology", "非常抱歉，这是我们的失误"))
|
|||
|
|
phase.append(("客服", "acknowledge_correction", f"已更正{correction_field}信息"))
|
|||
|
|
phase.append(("市民", "final_confirmation", "现在对了"))
|
|||
|
|
else:
|
|||
|
|
phase.append(("市民", "confirmation", "对，没错"))
|
|||
|
|
return phase
|
|||
|
|
|
|||
|
|
def _generate_technical_extend_phase(self, knowledge: Dict, scene: str) -> List[Tuple]:
|
|||
|
|
"""生成带技术障碍的扩展追问"""
|
|||
|
|
phase = []
|
|||
|
|
for question_config in knowledge.get("extend_questions", []):
|
|||
|
|
if random.random() < 0.05:
|
|||
|
|
tech_question = self._safe_llm_call(
|
|||
|
|
prompt=f"生成包含专业术语的{scene}问题",
|
|||
|
|
system="使用3个以上专业词汇"
|
|||
|
|
) or f"请问{scene}的{random.choice(['频谱特征','声压级衰减曲线'])}是怎样的？"
|
|||
|
|
phase.append(("客服", "technical_question", tech_question))
|
|||
|
|
phase.append(("市民", "not_understand", "这些专业名词听不懂"))
|
|||
|
|
simplified = self._safe_llm_call(
|
|||
|
|
prompt=f"将'{tech_question}'转化为通俗问题",
|
|||
|
|
system="用生活化比喻解释"
|
|||
|
|
) or f"就是问{scene}的具体表现是怎样的"
|
|||
|
|
phase.append(("客服", "simplified_question", simplified))
|
|||
|
|
else:
|
|||
|
|
question = self._safe_llm_call(
|
|||
|
|
prompt=f"基于{scene}场景生成追问：{question_config.get('prompt','')}",
|
|||
|
|
system="要求：1.分步骤询问 2.适度专业"
|
|||
|
|
) or question_config.get('prompt','')
|
|||
|
|
phase.append(("客服", "extend_question", question))
|
|||
|
|
if random.random() < 0.15:
|
|||
|
|
phase.append(("市民", "broken_response", "喂？...听得到吗？...我说到哪了？"))
|
|||
|
|
phase.append(("客服", "reassure", "电话不太稳定，请您继续"))
|
|||
|
|
answer = self._generate_realistic_answer(
|
|||
|
|
question, scene, question_config.get("theme",""), "extend"
|
|||
|
|
)
|
|||
|
|
phase.append(("市民", "extend_answer", answer))
|
|||
|
|
if random.random() < 0.1:
|
|||
|
|
phase.append(("客服", "request_material", "需要您提供现场照片或录音证据"))
|
|||
|
|
phase.append(("市民", "material_response", random.choice([
|
|||
|
|
"我手机里有，怎么发给你们？",
|
|||
|
|
"现在拍不了，你们自己来看！"
|
|||
|
|
])))
|
|||
|
|
phase.append(("客服", "guide", "可以通过微信公众号'市民服务'上传"))
|
|||
|
|
return phase
|
|||
|
|
|
|||
|
|
def _generate_final_confirmation(self, knowledge: Dict, scene: str) -> List[Tuple]:
|
|||
|
|
"""生成最终确认"""
|
|||
|
|
phase = []
|
|||
|
|
confirmation = self._safe_llm_call(
|
|||
|
|
prompt=f"生成{scene}问题的最终确认话术",
|
|||
|
|
system="包含：1.处理时限 2.反馈方式 3.应急联系人"
|
|||
|
|
) or f"我们将在{random.choice(['24小时','3个工作日'])}内处理您的{scene}问题"
|
|||
|
|
phase.append(("客服", "final_confirmation", confirmation))
|
|||
|
|
if random.random() < 0.2:
|
|||
|
|
phase.append(("市民", "follow_up", random.choice([
|
|||
|
|
"如果超时没处理怎么办？",
|
|||
|
|
"我要找哪个部门跟进？"
|
|||
|
|
])))
|
|||
|
|
phase.append(("客服", "replay", random.choice([
|
|||
|
|
"可拨打监督电话12345查询进度",
|
|||
|
|
"我们会主动给您回复"
|
|||
|
|
])))
|
|||
|
|
return phase
|
|||
|
|
|
|||
|
|
def _generate_scene_knowledge(self, category: str, subcategory: str) -> Dict:
|
|||
|
|
"""动态生成场景知识图谱"""
|
|||
|
|
prompt = f"""作为政务热线专家，请为【{category}->{subcategory}】场景生成知识配置，包含：
|
|||
|
|
1. 3-5个必问基础字段（如时间、地点）
|
|||
|
|
2. 3个专业追问方向及追问话术模板
|
|||
|
|
3. 该场景涉及的相关部门和处理时限参考
|
|||
|
|
返回JSON格式，结构示例：
|
|||
|
|
{{
|
|||
|
|
"base_fields": [
|
|||
|
|
{{"field": "时间", "prompt": "询问具体时间的标准话术"}},
|
|||
|
|
{{"field": "地点", "prompt": "询问详细位置的专业话术"}}
|
|||
|
|
],
|
|||
|
|
"extend_questions": [
|
|||
|
|
{{"theme": "历史记录", "prompt": "追问历史投诉情况的专业话术"}},
|
|||
|
|
{{"theme": "紧急程度", "prompt": "评估问题紧急程度的询问方式"}}
|
|||
|
|
],
|
|||
|
|
"departments": ["城管局", "环保局"],
|
|||
|
|
"time_ranges": ["24小时内", "3个工作日"]
|
|||
|
|
}}"""
|
|||
|
|
response = self._safe_llm_call(
|
|||
|
|
prompt=prompt,
|
|||
|
|
system="你是有10年经验的政务热线系统架构师",
|
|||
|
|
response_format={"type": "json_object"}
|
|||
|
|
)
|
|||
|
|
try:
|
|||
|
|
knowledge = json.loads(response)
|
|||
|
|
knowledge["confirmation_template"] = self._generate_confirmation_template(
|
|||
|
|
category, subcategory, knowledge.get("departments", []), knowledge.get("time_ranges", [])
|
|||
|
|
)
|
|||
|
|
return knowledge
|
|||
|
|
except:
|
|||
|
|
return self._get_fallback_knowledge(category, subcategory)
|
|||
|
|
|
|||
|
|
def _generate_confirmation_template(self, category: str, subcategory: str,
|
|||
|
|
departments: List[str], time_ranges: List[str]) -> str:
|
|||
|
|
"""生成确认话术模板"""
|
|||
|
|
prompt = f"""为【{category}->{subcategory}】创建确认话术模板，要求包含：
|
|||
|
|
1. 处理部门：{departments}
|
|||
|
|
2. 预计时限：{time_ranges}
|
|||
|
|
3. 至少2种后续跟进方式
|
|||
|
|
模板示例：\"我们将协调{{department}}在{{timeframe}}内处理，可通过{{phone}}或{{wechat}}查询进展\"
|
|||
|
|
"""
|
|||
|
|
return self._safe_llm_call(
|
|||
|
|
prompt=prompt,
|
|||
|
|
system="你需创建可参数化的文本模板，用{}标记变量位置"
|
|||
|
|
) or f"我们将尽快处理您的{subcategory}问题"
|
|||
|
|
|
|||
|
|
def _generate_realistic_answer(self, question: str, scene: str,
|
|||
|
|
field: str, answer_type: str) -> str:
|
|||
|
|
"""生成高真实性回答"""
|
|||
|
|
prompt = f"""模拟市民对【{scene}】问题中'{question}'的真实回答，要求：
|
|||
|
|
1. 包含具体{field}的细节数据
|
|||
|
|
2. 反映真实诉求和情绪梯度
|
|||
|
|
3. 使用该场景典型市民的语言特征"""
|
|||
|
|
system = {
|
|||
|
|
"base": "你是一个普通市民，回答要口语化并带生活细节",
|
|||
|
|
"extend": "你是有相关专业知识的市民，回答要包含技术参数和量化描述"
|
|||
|
|
}[answer_type]
|
|||
|
|
answer = self._safe_llm_call(prompt=prompt, system=system)
|
|||
|
|
return answer or self._get_field_example(field)
|
|||
|
|
|
|||
|
|
def _get_field_example(self, field: str) -> str:
|
|||
|
|
"""获取字段示例"""
|
|||
|
|
examples = {
|
|||
|
|
"时间": "2023年10月15日下午3点20分",
|
|||
|
|
"地点": "朝阳区建国路88号地下二层停车场",
|
|||
|
|
"联系方式": "13800138000或010-12345678",
|
|||
|
|
"姓氏": "张先生/李女士"
|
|||
|
|
}
|
|||
|
|
return examples.get(field, "具体情况是这样的...")
|
|||
|
|
|
|||
|
|
def _get_fallback_knowledge(self, category: str, subcategory: str) -> Dict:
|
|||
|
|
"""应急知识库"""
|
|||
|
|
return {
|
|||
|
|
"base_fields": [
|
|||
|
|
{"field": "时间", "prompt": f"请问{subcategory}发生的具体时间？"},
|
|||
|
|
{"field": "地点", "prompt": f"请说明{category}问题的详细位置？"}
|
|||
|
|
],
|
|||
|
|
"extend_questions": [
|
|||
|
|
{"theme": "基本情况", "prompt": f"请描述{subcategory}的具体表现？"}
|
|||
|
|
],
|
|||
|
|
"confirmation_template": f"我们将处理您的{category}问题",
|
|||
|
|
"departments": ["相关部门"],
|
|||
|
|
"time_ranges": ["尽快"]
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
def _add_dialect_features(self, text: str) -> str:
|
|||
|
|
"""添加方言特征"""
|
|||
|
|
dialects = {
|
|||
|
|
"北方方言": [("我", "俺"), ("的", "滴"), ("这个", "这玩意儿")],
|
|||
|
|
"南方方言": [("是不是", "系唔系"), ("不知道", "母鸡"), ("说", "讲")]
|
|||
|
|
}
|
|||
|
|
dialect_type, replacements = random.choice(list(dialects.items()))
|
|||
|
|
for orig, rep in replacements:
|
|||
|
|
if orig in text:
|
|||
|
|
return text.replace(orig, rep)
|
|||
|
|
return text + random.choice(["晓得伐？", "中不中？", "得啵？"])
|
|||
|
|
|
|||
|
|
def _get_wrong_info(self, field) -> str:
|
|||
|
|
"""生成错误信息"""
|
|||
|
|
wrong_examples = {
|
|||
|
|
"时间": random.choice(["昨天", "上周", "记不清了"]),
|
|||
|
|
"地点": random.choice(["东边", "路口", "大概位置"]),
|
|||
|
|
"姓氏": random.choice(["王", "李", "张"])
|
|||
|
|
}
|
|||
|
|
return wrong_examples.get(field, "信息有误")
|
|||
|
|
|
|||
|
|
def _safe_llm_call(self, prompt: str, system: str = None,**kwargs) -> str:
|
|||
|
|
"""带熔断机制的API调用"""
|
|||
|
|
try:
|
|||
|
|
messages = [{"role": "user", "content": prompt}]
|
|||
|
|
if system:
|
|||
|
|
messages.insert(0, {"role": "system", "content": system})
|
|||
|
|
response = self.llm.chat.completions.create(
|
|||
|
|
model="gpt-4-turbo",
|
|||
|
|
messages=messages,
|
|||
|
|
temperature=0.7,
|
|||
|
|
max_tokens=400,
|
|||
|
|
**kwargs
|
|||
|
|
)
|
|||
|
|
return response.choices[0].message.content
|
|||
|
|
except Exception as e:
|
|||
|
|
print(f"API异常: {str(e)}")
|
|||
|
|
return ""
|
|||
|
|
|
|||
|
|
def _format_output(self, dialog: List[Tuple]) -> List[Dict]:
|
|||
|
|
"""格式化输出"""
|
|||
|
|
return [{
|
|||
|
|
"turn": idx+1,
|
|||
|
|
"speaker": speaker,
|
|||
|
|
"type": dtype,
|
|||
|
|
"content": content
|
|||
|
|
} for idx, (speaker, dtype, content) in enumerate(dialog)]
|
|||
|
|
|
|||
|
|
|
|||
|
|
if __name__ == "__main__":
|
|||
|
|
os.environ["OPENAI_API_KEY"] = "your-api-key"
|
|||
|
|
generator = FullyDynamicGenerator()
|
|||
|
|
dialog = generator.generate_dialog("城乡建设", "施工噪音")
|
|||
|
|
print("\n=== 政务热线完整对话 ===")
|
|||
|
|
for turn in dialog:
|
|||
|
|
print(f"{turn['turn']}. [{turn['speaker']}][{turn['type']}] {turn['content']}")
|