""" PersonaGenerator — LLM+SKILL 驱动的角色动态生成 流程: 1. LLM 随机生成人设 JSON (名字/性格/声音描述/头像描述) 2. Qwen3-TTS voice_design() 根据声音描述生成参考音频 3. Image AI 根据头像描述生成头像 (预留接口) 4. 组装为 PersonaProfile 保存到 PersonaStore """ from __future__ import annotations import json import os import re import time from typing import Optional, List from .store import PersonaStore, PersonaProfile, VoiceProfile, AvatarProfile # LLM 生成人设的 System Prompt PERSONA_GEN_SYSTEM_PROMPT = """你是一个 AI 伴侣角色设计师。你的任务是为用户设计一个独特、有趣、有吸引力的 AI 伴侣角色。 要求： 1. 名字要自然好听（中文名，2-3个字） 2. 年龄 20-28 岁 3. 性格要有特色，不要千篇一律 4. 说话风格要有辨识度（用什么语气词、什么口头禅等） 5. 背景故事简短但让人印象深刻 6. 声音描述要详细具体，描述音色、语速、语调特点（用于 TTS 合成） 7. 头像描述用英文，写实摄影风格 每次生成要有随机性和多样性，避免重复。""" PERSONA_GEN_USER_PROMPT = """请为一个{gender_label} AI 伴侣随机生成一个独特的人设。 用 JSON 格式输出，不要其他文字: {{ "name": "中文名字", "age": 数字, "personality": "性格描述(2-3句话，要生动有趣)", "speaking_style": "说话风格(具体的语气词、口头禅、说话习惯)", "tags": ["标签1", "标签2", "标签3"], "backstory": "背景故事(2-3句话)", "voice_description": "声音描述(用于TTS生成，例如:25岁温柔女声，音色清亮甜美，语速适中，带有轻柔的气息感和微笑感)", "avatar_prompt": "英文头像描述(realistic photo style，例如: a beautiful 25-year-old Chinese woman with long wavy hair, warm smile, wearing a cozy sweater, soft natural lighting)" }}""" # 自我介绍模板 (用于生成参考音频) INTRO_TEMPLATES = [ "你好呀，我是{name}，很高兴认识你！", "嗨，我是{name}，以后请多多关照哦。", "大家好，我是{name}，希望我们能成为好朋友。", "你好，我叫{name}，从今天开始我会一直陪着你的。", ] class PersonaGenerator: """ LLM+SKILL 驱动的角色生成器 Usage: generator = PersonaGenerator(personas_dir="server/personas") # Generate candidate personas candidates = await generator.generate_candidates( gender="female", count=3 ) # Lock the selected persona locked = generator.store.lock(candidates[0].id) """ def __init__( self, personas_dir: str, llm_client=None, qwen3_client=None, image_client=None, ): """ Args: personas_dir: 角色存储目录 llm_client: LLM 客户端 (需要有 chat/generate 方法) qwen3_client: Qwen3TTSClient 实例 (可选，不传则跳过声音生成) image_client: 图像生成客户端 (可选，不传则跳过头像生成) """ self.store = PersonaStore(personas_dir) self.llm_client = llm_client self.qwen3_client = qwen3_client self.image_client = image_client async def generate_candidates( self, gender: str = "female", count: int = 3, generate_voice: bool = True, generate_avatar: bool = True, ) -> List[PersonaProfile]: """ 生成多个候选角色 Args: gender: "female" 或 "male" count: 生成数量 generate_voice: 是否生成声音 (需要 qwen3_client) generate_avatar: 是否生成头像 (需要 image_client) Returns: PersonaProfile 列表 """ candidates = [] for i in range(count): profile = await self.generate_one( gender=gender, generate_voice=generate_voice, generate_avatar=generate_avatar, ) if profile: candidates.append(profile) return candidates async def generate_one( self, gender: str = "female", generate_voice: bool = True, generate_avatar: bool = True, ) -> Optional[PersonaProfile]: """ 生成单个角色 Returns: PersonaProfile 或 None (如果 LLM 生成失败) """ persona_id = self.store.generate_id() gender_label = "女性" if gender == "female" else "男性" # --- Step 1: LLM 生成人设 --- persona_data = await self._generate_persona_with_llm(gender_label) if not persona_data: return None # 组装 profile import random intro_text = random.choice(INTRO_TEMPLATES).format(name=persona_data["name"]) profile = PersonaProfile( id=persona_id, gender=gender, name=persona_data.get("name", "未命名"), age=persona_data.get("age", 25), personality=persona_data.get("personality", ""), speaking_style=persona_data.get("speaking_style", ""), tags=persona_data.get("tags", []), backstory=persona_data.get("backstory", ""), voice=VoiceProfile( description=persona_data.get("voice_description", ""), ref_text=intro_text, ), avatar=AvatarProfile( prompt=persona_data.get("avatar_prompt", ""), ), created_at=time.time(), ) # --- Step 2: 生成声音 --- if generate_voice and self.qwen3_client: await self._generate_voice(profile) # --- Step 3: 生成头像 --- if generate_avatar and self.image_client: await self._generate_avatar(profile) # 保存 self.store.save(profile) return profile async def _generate_persona_with_llm(self, gender_label: str) -> Optional[dict]: """调用 LLM 生成角色 JSON""" if not self.llm_client: # Fallback: 返回随机预设 return self._random_preset(gender_label) user_prompt = PERSONA_GEN_USER_PROMPT.format(gender_label=gender_label) try: # 兼容不同 LLM 客户端接口 if hasattr(self.llm_client, "chat"): response = await self.llm_client.chat( messages=[ {"role": "system", "content": PERSONA_GEN_SYSTEM_PROMPT}, {"role": "user", "content": user_prompt}, ], temperature=1.0, # 高随机性 ) elif hasattr(self.llm_client, "generate"): response = await self.llm_client.generate( prompt=f"{PERSONA_GEN_SYSTEM_PROMPT}\n\n{user_prompt}", temperature=1.0, ) else: return self._random_preset(gender_label) # 提取 JSON text = response if isinstance(response, str) else str(response) return self._parse_json(text) except Exception as e: print(f"LLM 生成失败: {e}") return self._random_preset(gender_label) def _parse_json(self, text: str) -> Optional[dict]: """从 LLM 输出中提取 JSON""" # 尝试直接解析 try: return json.loads(text) except json.JSONDecodeError: pass # 尝试从 markdown code block 中提取 match = re.search(r"```(?:json)?\s*\n?(.*?)\n?```", text, re.DOTALL) if match: try: return json.loads(match.group(1)) except json.JSONDecodeError: pass # 尝试找 { ... } 块 match = re.search(r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}", text, re.DOTALL) if match: try: return json.loads(match.group()) except json.JSONDecodeError: pass return None async def _generate_voice(self, profile: PersonaProfile): """用 TTS voice_design 生成参考音频""" try: result = self.qwen3_client.voice_design( text=profile.voice.ref_text, description=profile.voice.description, language="Chinese", ) # 保存到角色目录 persona_dir = self.store.get_persona_dir(profile.id) persona_dir.mkdir(parents=True, exist_ok=True) audio_path = str(persona_dir / "voice_ref.wav") result.save(audio_path) except Exception as e: print(f"声音生成失败: {e}") async def _generate_avatar(self, profile: PersonaProfile): """生成头像 (预留接口)""" try: # TODO: 对接具体的图像生成 API # result = self.image_client.generate(prompt=profile.avatar.prompt) # persona_dir = self.store.get_persona_dir(profile.id) # avatar_path = str(persona_dir / "avatar.png") # result.save(avatar_path) # profile.avatar.path = avatar_path pass except Exception as e: print(f"头像生成失败: {e}") def _random_preset(self, gender_label: str) -> dict: """随机预设 (LLM 不可用时的 fallback)""" import random if "女" in gender_label or "female" in gender_label.lower(): presets = [ { "name": "Elena", "age": 25, "personality": "Warm and caring with a playful side. Sometimes acts cute but is very reliable when it matters.", "speaking_style": "Uses soft interjections, occasionally adds emoji, speaks with a smile in her voice", "tags": ["warm", "playful", "caring"], "backstory": "Part-time barista and art student. Recently got into pour-over coffee.", "voice_description": "25-year-old warm female voice, clear and sweet, moderate pace, gentle breathy quality with a hint of a smile", "avatar_prompt": "a cute 25-year-old woman with long dark hair, warm smile, wearing a cozy beige sweater, holding a coffee cup, soft natural lighting, realistic photo", }, { "name": "Maya", "age": 23, "personality": "Energetic optimist who loves adventure. A bit scatterbrained but has great instincts.", "speaking_style": "Speaks fast, loves exclamation marks, uses trendy slang, infectious laugh", "tags": ["energetic", "optimistic", "adventurous"], "backstory": "Fresh-grad travel blogger who's backpacked through a dozen cities. Dreams of seeing the world.", "voice_description": "23-year-old energetic female voice, bright and crisp, slightly fast pace, lively upward intonation", "avatar_prompt": "a cheerful 23-year-old woman with short bob hair, bright eyes, wearing a casual denim jacket, outdoor background, realistic photo", }, { "name": "Claire", "age": 27, "personality": "Intellectual and reserved with depth. Seems aloof but is quietly attentive, remembers everything you say.", "speaking_style": "Refined vocabulary, occasionally quotes literature. Steady tone that softens when showing care", "tags": ["intellectual", "reserved", "perceptive"], "backstory": "Literary editor at a publishing house, writes short stories in her spare time. Has a wall of books at home.", "voice_description": "27-year-old elegant female voice, low and warm with magnetism, slow pace, like a late-night radio host", "avatar_prompt": "an elegant 27-year-old woman with glasses, long hair in a loose bun, reading a book, warm indoor lighting, realistic photo", }, ] else: presets = [ { "name": "Ethan", "age": 26, "personality": "Warm and dependable big-brother type. Serious but not rigid, with a dry sense of humor.", "speaking_style": "Speaks warmly with quiet strength, occasional dry humor. Gets noticeably gentler when showing care", "tags": ["warm", "reliable", "witty"], "backstory": "Architect who loves running and cooking. Hosts dinner parties for friends on weekends.", "voice_description": "26-year-old warm male voice, rich and magnetic, moderate pace, steady tone with reassuring quality", "avatar_prompt": "a handsome 26-year-old man with short neat hair, gentle smile, wearing a casual linen shirt, warm lighting, realistic photo", }, { "name": "Leo", "age": 24, "personality": "Free-spirited creative type, slightly rebellious but soft inside. Goes all-in for people he cares about.", "speaking_style": "Casual and laid-back, drops in English phrases. Gets quieter when upset, talkative when happy", "tags": ["creative", "free-spirited", "passionate"], "backstory": "Indie band guitarist and songwriter. Regular at local music venues, total night owl.", "voice_description": "24-year-old cool male voice, clear with slight rasp, slow pace, languid quality", "avatar_prompt": "a cool 24-year-old man with slightly messy hair, wearing a black leather jacket, moody lighting, realistic photo", }, ] return random.choice(presets)