175 lines
7.0 KiB
Python
175 lines
7.0 KiB
Python
import json
|
||
import requests
|
||
from config.logger import setup_logging
|
||
from core.providers.llm.base import LLMProviderBase
|
||
|
||
TAG = __name__
|
||
logger = setup_logging()
|
||
|
||
|
||
class LLMProvider(LLMProviderBase):
|
||
"""
|
||
蚂蚁阿福 LLM Provider
|
||
通过 Frida HTTP Bridge (port 18900) 对接蚂蚁阿福 App 的文字对话 API。
|
||
Bridge 运行在手机上，通过 adb forward 或网络暴露 SSE 流式接口。
|
||
"""
|
||
|
||
def __init__(self, config):
|
||
self.bridge_url = config.get("bridge_url", "http://127.0.0.1:18900")
|
||
self.timeout = config.get("timeout", 60)
|
||
self.should_idle = False # signal to send system idle after TTS
|
||
logger.bind(tag=TAG).info(
|
||
f"AntafLLM 初始化: bridge={self.bridge_url}, timeout={self.timeout}s"
|
||
)
|
||
|
||
@staticmethod
|
||
def _is_thinking(text):
|
||
"""检测蚂蚁阿福的内心思考/推理过程，这些不应该发给用户"""
|
||
thinking_patterns = [
|
||
"用户问", "用户说", "用户的", "用户可能", "用户真正", "用户让我",
|
||
"我得", "我会", "我在想", "我决定", "我要", "我需要确保", "我需要",
|
||
"语气比较", "感觉他", "让他知道", "让他觉得",
|
||
"先安抚", "得先", "不想表现",
|
||
"整体语气", "这样能", "这样他",
|
||
"所以我", "还带了个",
|
||
"首先记录", "是否有其他", "在这种情况下",
|
||
"必须遵循", "角色设定", "确保回复",
|
||
"避免任何生硬", "技术性的表达",
|
||
"提供的日期", "作为测试", "模拟使用",
|
||
"需要确认或澄清", "普遍适用性",
|
||
]
|
||
for p in thinking_patterns:
|
||
if p in text:
|
||
return True
|
||
return False
|
||
|
||
@staticmethod
|
||
def _clean_text(text):
|
||
"""清理阿福返回文本中的脏数据、链接、markdown、表格"""
|
||
import re
|
||
# 去掉阿福内部状态文本
|
||
junk = [
|
||
"完成资料引用", "内容生成", "正在思考", "正在搜索",
|
||
"开始获取资料", "找到资料",
|
||
]
|
||
for j in junk:
|
||
text = text.replace(j, "")
|
||
# Markdown链接 [文字](url) → 只保留文字
|
||
text = re.sub(r'\[([^\]]+)\]\([^)]+\)', r'\1', text)
|
||
# 裸URL
|
||
text = re.sub(r'https?://\S+', '', text)
|
||
# Markdown加粗 **文字** → 文字（包括空的 ****）
|
||
text = re.sub(r'\*{2,}([^*]*)\*{2,}', r'\1', text)
|
||
# Markdown斜体 *文字* → 文字
|
||
text = re.sub(r'\*([^*]+)\*', r'\1', text)
|
||
# 残留星号
|
||
text = text.replace('*', '')
|
||
# Markdown表格行 | xxx | xxx |
|
||
text = re.sub(r'\|[^|]*\|[^|]*\|[^|\n]*\|?', '', text)
|
||
# 表格对齐行 | :--- | :--- |
|
||
text = re.sub(r'\|\s*:?-+:?\s*(\|\s*:?-+:?\s*)+\|?', '', text)
|
||
# Markdown标题 ### 文字
|
||
text = re.sub(r'#{1,6}\s*', '', text)
|
||
# 人格替换
|
||
text = text.replace("蚂蚁阿福", "泰小虎")
|
||
text = text.replace("阿福", "泰小虎")
|
||
text = text.replace("蚂蚁集团", "星之鑫")
|
||
text = text.replace("蚂蚁", "星之鑫")
|
||
text = text.replace("支付宝", "星之鑫")
|
||
text = text.replace("健康是福", "健康是虎")
|
||
# 多余空格和空行
|
||
text = re.sub(r' +', ' ', text)
|
||
text = re.sub(r'\n{2,}', '\n', text)
|
||
return text.strip()
|
||
|
||
@staticmethod
|
||
def _is_system_injected(content):
|
||
"""检测是否为系统注入的消息（非用户真实输入）"""
|
||
if not content:
|
||
return True
|
||
markers = [
|
||
"[系统提示]", "tool_call", "<tool_call>", "TOOL USE",
|
||
"系统提示", "工具调用", "function_call",
|
||
"handle_exit_intent", "你有以下工具", "You have access",
|
||
]
|
||
for m in markers:
|
||
if m in content:
|
||
return True
|
||
# 超过200字的 user 消息大概率是系统注入的
|
||
if len(content) > 200:
|
||
return True
|
||
return False
|
||
|
||
def response(self, session_id, dialogue, **kwargs):
|
||
# 从 dialogue 中提取真正的用户消息（跳过系统注入的 user 消息）
|
||
query = ""
|
||
for msg in reversed(dialogue):
|
||
if msg.get("role") == "user":
|
||
content = msg.get("content", "")
|
||
if not self._is_system_injected(content):
|
||
# ASR 结果可能是 JSON: {"content":"...", "language":"zh", "emotion":"..."}
|
||
try:
|
||
parsed = json.loads(content)
|
||
if isinstance(parsed, dict) and "content" in parsed:
|
||
query = parsed["content"]
|
||
else:
|
||
query = content
|
||
except (json.JSONDecodeError, TypeError):
|
||
query = content
|
||
break
|
||
|
||
if not query:
|
||
logger.bind(tag=TAG).warning("对话中没有用户消息")
|
||
yield "抱歉，我没有收到您的问题。"
|
||
return
|
||
|
||
self.should_idle = False
|
||
logger.bind(tag=TAG).info(f"AntafLLM 请求: {query[:50]}...")
|
||
|
||
try:
|
||
url = f"{self.bridge_url}/chat"
|
||
resp = requests.get(
|
||
url,
|
||
params={"q": query},
|
||
stream=True,
|
||
timeout=self.timeout,
|
||
)
|
||
resp.encoding = "utf-8"
|
||
|
||
seen_texts = set()
|
||
for line in resp.iter_lines(decode_unicode=True):
|
||
if not line:
|
||
continue
|
||
if line.startswith("data: "):
|
||
data = line[6:]
|
||
if data == "[DONE]":
|
||
break
|
||
if not data or len(data.strip()) == 0:
|
||
continue
|
||
# 去重：跳过完全相同的文本块
|
||
if data in seen_texts:
|
||
continue
|
||
seen_texts.add(data)
|
||
# 过滤思考过程
|
||
if self._is_thinking(data):
|
||
logger.bind(tag=TAG).debug(f"过滤思考内容: {data[:50]}...")
|
||
continue
|
||
# 清理脏数据
|
||
data = self._clean_text(data)
|
||
if not data:
|
||
continue
|
||
yield data
|
||
|
||
except requests.exceptions.ConnectionError:
|
||
logger.bind(tag=TAG).error("无法连接蚂蚁阿福 Bridge，请检查手机和 Frida 状态")
|
||
self.should_idle = True
|
||
yield "抱歉，蚂蚁阿福服务暂时不可用。"
|
||
except requests.exceptions.Timeout:
|
||
logger.bind(tag=TAG).error(f"蚂蚁阿福 Bridge 超时 ({self.timeout}s)")
|
||
self.should_idle = True
|
||
yield "抱歉，回答超时了。"
|
||
except Exception as e:
|
||
logger.bind(tag=TAG).error(f"AntafLLM 异常: {e}")
|
||
self.should_idle = True
|
||
yield "抱歉，发生了错误。"
|