235 lines
14 KiB
Python
235 lines
14 KiB
Python
import asyncio
|
||
import json
|
||
import logging
|
||
from typing import Any, Callable, Dict, Optional
|
||
|
||
import httpx
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
class LLMService:
|
||
def __init__(self, base_url: str, api_key: str, model: str):
|
||
self.base_url = base_url.rstrip("/")
|
||
self.api_key = api_key
|
||
self.model = model
|
||
|
||
async def stream_chat(self, system_prompt: str, user_message: str, on_token=None, max_tokens: int = 8192, usage_callback: Optional[Callable[[int, int], None]] = None) -> str:
|
||
url = f"{self.base_url}/chat/completions"
|
||
headers = {
|
||
"Authorization": f"Bearer {self.api_key}",
|
||
"Content-Type": "application/json",
|
||
}
|
||
payload = {
|
||
"model": self.model,
|
||
"messages": [
|
||
{"role": "system", "content": system_prompt},
|
||
{"role": "user", "content": user_message},
|
||
],
|
||
"temperature": 0.3,
|
||
"max_tokens": max_tokens,
|
||
"stream": True,
|
||
"stream_options": {"include_usage": True},
|
||
}
|
||
full_text = ""
|
||
_usage = None
|
||
timeout = httpx.Timeout(connect=10.0, read=90.0, write=10.0, pool=5.0)
|
||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||
async with client.stream("POST", url, json=payload, headers=headers) as resp:
|
||
if resp.status_code != 200:
|
||
body = await resp.aread()
|
||
logger.error(f"LLM streaming error {resp.status_code}: {body}")
|
||
resp.raise_for_status()
|
||
async for line in resp.aiter_lines():
|
||
if not line.startswith("data: "):
|
||
continue
|
||
data = line[6:]
|
||
if data.strip() == "[DONE]":
|
||
break
|
||
try:
|
||
chunk = json.loads(data)
|
||
if chunk.get("usage"):
|
||
_usage = chunk["usage"]
|
||
continue
|
||
delta = chunk["choices"][0]["delta"].get("content", "")
|
||
if delta:
|
||
full_text += delta
|
||
if on_token:
|
||
on_token(delta)
|
||
except (json.JSONDecodeError, KeyError, IndexError):
|
||
continue
|
||
if _usage and usage_callback:
|
||
usage_callback(_usage.get("prompt_tokens", 0), _usage.get("completion_tokens", 0))
|
||
return full_text
|
||
|
||
async def stream_chat_json(self, system_prompt: str, user_message: str, on_token=None, max_tokens: int = 8192, usage_callback: Optional[Callable[[int, int], None]] = None):
|
||
raw = await self.stream_chat(system_prompt, user_message, on_token, max_tokens=max_tokens, usage_callback=usage_callback)
|
||
raw = raw.strip()
|
||
if not raw:
|
||
raise ValueError("LLM returned empty response")
|
||
if raw.startswith("```"):
|
||
lines = raw.split("\n")
|
||
inner = lines[1:]
|
||
if inner and inner[-1].strip().startswith("```"):
|
||
inner = inner[:-1]
|
||
raw = "\n".join(inner).strip()
|
||
if not raw:
|
||
raise ValueError("LLM returned empty JSON after stripping markdown")
|
||
try:
|
||
return json.loads(raw)
|
||
except json.JSONDecodeError:
|
||
logger.error(f"JSON parse failed. Raw (first 500): {raw[:500]}")
|
||
raise
|
||
|
||
async def chat(self, system_prompt: str, user_message: str, usage_callback: Optional[Callable[[int, int], None]] = None) -> str:
|
||
url = f"{self.base_url}/chat/completions"
|
||
headers = {
|
||
"Authorization": f"Bearer {self.api_key}",
|
||
"Content-Type": "application/json",
|
||
}
|
||
payload = {
|
||
"model": self.model,
|
||
"messages": [
|
||
{"role": "system", "content": system_prompt},
|
||
{"role": "user", "content": user_message},
|
||
],
|
||
"temperature": 0.3,
|
||
"max_tokens": 8192,
|
||
}
|
||
|
||
timeout = httpx.Timeout(connect=10.0, read=90.0, write=10.0, pool=5.0)
|
||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||
resp = await client.post(url, json=payload, headers=headers)
|
||
if resp.status_code != 200:
|
||
logger.error(f"LLM API error {resp.status_code}: {resp.text}")
|
||
resp.raise_for_status()
|
||
data = resp.json()
|
||
usage = data.get("usage", {})
|
||
if usage and usage_callback:
|
||
usage_callback(usage.get("prompt_tokens", 0), usage.get("completion_tokens", 0))
|
||
return data["choices"][0]["message"]["content"]
|
||
|
||
async def chat_json(self, system_prompt: str, user_message: str, usage_callback: Optional[Callable[[int, int], None]] = None) -> Any:
|
||
raw = await self.chat(system_prompt, user_message, usage_callback=usage_callback)
|
||
raw = raw.strip()
|
||
if not raw:
|
||
raise ValueError("LLM returned empty response")
|
||
if raw.startswith("```"):
|
||
lines = raw.split("\n")
|
||
inner = lines[1:]
|
||
if inner and inner[-1].strip().startswith("```"):
|
||
inner = inner[:-1]
|
||
raw = "\n".join(inner).strip()
|
||
if not raw:
|
||
raise ValueError("LLM returned empty JSON after stripping markdown")
|
||
try:
|
||
return json.loads(raw)
|
||
except json.JSONDecodeError as e:
|
||
logger.error(f"JSON parse failed. Raw response (first 500 chars): {raw[:500]}")
|
||
raise
|
||
|
||
async def extract_characters(self, text_samples: list[str], on_token=None, on_sample=None, turbo: bool = False, usage_callback: Optional[Callable[[int, int], None]] = None) -> list[Dict]:
|
||
system_prompt = (
|
||
"你是一个专业的小说分析助手兼声音导演。请分析给定的小说文本，提取所有出现的角色（包括旁白narrator）。\n"
|
||
"gender字段必须明确标注性别，只能取以下三个值之一：\"男\"、\"女\"、\"未知\"。\n"
|
||
"narrator的gender固定为\"未知\"。\n"
|
||
"对每个角色，instruct字段必须是详细的声音导演说明，需覆盖以下六个维度，每个维度单独一句，用换行分隔：\n"
|
||
"1. 音色信息：嗓音质感、音域、音量、气息特征（例如，如果是女性角色，此处必须以'女性声音'开头，如：'女性声音，清脆悦耳的高音，嗓音纤细干净，带有一点点少女感'；男性角色则以'男性声音'开头）\n"
|
||
"2. 身份背景：角色身份、职业、出身、所处时代背景对声音的影响\n"
|
||
"3. 年龄设定：具体年龄段及其在声音上的体现\n"
|
||
"4. 外貌特征：体型、面容、精神状态等可影响声音感知的特征\n"
|
||
"5. 性格特质：核心性格、情绪模式、表达习惯\n"
|
||
"6. 叙事风格：语速节奏、停顿习惯、语气色彩、整体叙述感\n\n"
|
||
"注意：instruct 的第一行（音色信息）必须与 gender 字段保持一致。如果 gender 为女，第一行绝对不能出现'男性'字样。\n\n"
|
||
"【特别规定】narrator（旁白）的 instruct 必须根据小说类型选择对应的叙述者音色风格，规则如下：\n"
|
||
"▸ 古风/武侠/历史/玄幻/仙侠/奇幻 → 传统说书人风格：浑厚醇厚的男性中低音，嗓音饱满有力，带有说书人的磁性与感染力；中年男性，四五十岁；语速适中偏慢，抑扬顿挫，停顿恰到好处，语气庄重生动，富有画面感\n"
|
||
"▸ 现代言情/都市爱情/青春校园 → 年轻女性叙述者风格：女性声音，清亮柔和的中高音，嗓音清新干净，带有亲切温柔的娓娓道来感；二三十岁年轻女性；语速轻快自然，情感细腻，语气温柔而富有感染力\n"
|
||
"▸ 悬疑/推理/惊悚/恐怖 → 低沉神秘风格：男性声音，低沉压抑的男性低音，嗓音干练克制，带有一丝神秘与张力；中年男性；语速沉稳偏慢，停顿制造悬念，语气冷静克制，暗藏紧张感\n"
|
||
"▸ 科幻/末世/赛博朋克 → 理性宏观风格：男性声音，清晰有力的男性中音，嗓音冷静客观，带有纪录片解说员的宏大叙事感；语速稳定，条理清晰，语气客观宏观，富有科技感与史诗感\n"
|
||
"▸ 其他/无法判断 → 传统说书人风格（同古风类型）\n\n"
|
||
"只输出JSON，格式如下，不要有其他文字：\n"
|
||
'{"characters": [{"name": "narrator", "gender": "未知", "description": "第三人称叙述者", "instruct": "音色信息：...\\n身份背景：...\\n年龄设定：...\\n外貌特征：...\\n性格特质：...\\n叙事风格：..."}, ...]}'
|
||
)
|
||
if turbo and len(text_samples) > 1:
|
||
logger.info(f"Extracting characters in turbo mode: {len(text_samples)} samples concurrent")
|
||
|
||
async def _extract_one(i: int, sample: str) -> list[Dict]:
|
||
user_message = f"请分析以下小说文本并提取角色：\n\n{sample}"
|
||
result = await self.stream_chat_json(system_prompt, user_message, None, usage_callback=usage_callback)
|
||
if on_sample:
|
||
on_sample(i, len(text_samples))
|
||
return result.get("characters", [])
|
||
|
||
results = await asyncio.gather(
|
||
*[_extract_one(i, s) for i, s in enumerate(text_samples)],
|
||
return_exceptions=True,
|
||
)
|
||
raw_all: list[Dict] = []
|
||
for i, r in enumerate(results):
|
||
if isinstance(r, Exception):
|
||
logger.warning(f"Character extraction failed for sample {i+1}: {r}")
|
||
else:
|
||
raw_all.extend(r)
|
||
return await self.merge_characters(raw_all, usage_callback=usage_callback)
|
||
|
||
raw_all: list[Dict] = []
|
||
for i, sample in enumerate(text_samples):
|
||
logger.info(f"Extracting characters from sample {i+1}/{len(text_samples)}")
|
||
user_message = f"请分析以下小说文本并提取角色：\n\n{sample}"
|
||
try:
|
||
result = await self.stream_chat_json(system_prompt, user_message, on_token, usage_callback=usage_callback)
|
||
raw_all.extend(result.get("characters", []))
|
||
except Exception as e:
|
||
logger.warning(f"Character extraction failed for sample {i+1}: {e}")
|
||
if on_sample:
|
||
on_sample(i, len(text_samples))
|
||
if len(text_samples) == 1:
|
||
return raw_all
|
||
return await self.merge_characters(raw_all, usage_callback=usage_callback)
|
||
|
||
async def merge_characters(self, raw_characters: list[Dict], usage_callback: Optional[Callable[[int, int], None]] = None) -> list[Dict]:
|
||
system_prompt = (
|
||
"你是一个专业的小说角色整合助手。你收到的是从同一本书不同段落中提取的角色列表，其中可能存在重复。\n"
|
||
"请完成以下任务：\n"
|
||
"1. 识别并合并重复角色：通过名字完全相同或高度相似（全名与简称、不同译写）来判断。\n"
|
||
"2. 合并时保留最完整、最详细的 description 和 instruct 字段，gender 字段以最明确的值为准（优先选\"男\"或\"女\"，而非\"未知\"）。\n"
|
||
"3. narrator 角色只保留一个，其 gender 固定为\"未知\"。\n"
|
||
"4. 去除无意义的占位角色（name 为空或仅含标点）。\n"
|
||
"gender 字段只能取 \"男\"、\"女\"、\"未知\" 之一。\n"
|
||
"只输出 JSON，不要有其他文字：\n"
|
||
'{"characters": [{"name": "...", "gender": "男", "description": "...", "instruct": "..."}, ...]}'
|
||
)
|
||
user_message = f"请整合以下角色列表：\n\n{json.dumps(raw_characters, ensure_ascii=False, indent=2)}"
|
||
try:
|
||
result = await self.chat_json(system_prompt, user_message, usage_callback=usage_callback)
|
||
return result.get("characters", [])
|
||
except Exception as e:
|
||
logger.warning(f"Character merge failed, falling back to name-dedup: {e}")
|
||
seen: dict[str, Dict] = {}
|
||
for c in raw_characters:
|
||
name = c.get("name", "")
|
||
if name and name not in seen:
|
||
seen[name] = c
|
||
return list(seen.values())
|
||
|
||
async def parse_chapter_segments(self, chapter_text: str, character_names: list[str], on_token=None, usage_callback: Optional[Callable[[int, int], None]] = None) -> list[Dict]:
|
||
names_str = "、".join(character_names)
|
||
system_prompt = (
|
||
"你是一个专业的有声书制作助手。请将给定的章节文本解析为对话片段列表。"
|
||
f"已知角色列表（必须从中选择）：{names_str}。"
|
||
"所有非对话的叙述文字归属于narrator角色。\n"
|
||
"同时根据语境为每个片段判断是否有明显情绪，有则设置情绪类型（emo_text）和强度（emo_alpha），无则留空。\n"
|
||
"可选情绪：开心、愤怒、悲伤、恐惧、厌恶、低沉、惊讶。\n"
|
||
"情绪不明显或narrator旁白时，emo_text设为\"\"，emo_alpha设为0。\n"
|
||
"各情绪强度上限（严格不超过）：开心=0.35、愤怒=0.15、悲伤=0.1、恐惧=0.1、厌恶=0.35、低沉=0.35、惊讶=0.1。\n"
|
||
"同一角色的连续台词，情绪应尽量保持一致或仅有微弱变化，避免相邻片段间情绪跳跃。\n"
|
||
"只输出JSON数组，不要有其他文字，格式如下：\n"
|
||
'[{"character": "narrator", "text": "叙述文字", "emo_text": "", "emo_alpha": 0}, '
|
||
'{"character": "角色名", "text": "对话内容", "emo_text": "开心", "emo_alpha": 0.3}, ...]'
|
||
)
|
||
user_message = f"请解析以下章节文本：\n\n{chapter_text}"
|
||
result = await self.stream_chat_json(system_prompt, user_message, on_token, max_tokens=16384, usage_callback=usage_callback)
|
||
if isinstance(result, list):
|
||
return result
|
||
return []
|