mirror of
https://github.com/xinnan-tech/xiaozhi-esp32-server.git
synced 2026-07-29 03:53:55 +08:00
本地记忆+意图识别 (#250)
* 增加本地记忆功能,使用llm总结记忆 * update:增加统一非流式输出输出 * 增加意图识别内容,使用llm进行识别 * 初始化记忆模块 * 完善意图识别处理后的流程 * 通过使用function call实现意图识别 * update:优化意图识别的配置 * update:function call最优设置成doubao-pro-32k-functioncall-241028 --------- Co-authored-by: 玄凤科技 <eric230308@gmail.com> Co-authored-by: hrz <1710360675@qq.com>
This commit is contained in:
@@ -0,0 +1,33 @@
|
||||
from abc import ABC, abstractmethod
|
||||
from typing import List, Dict
|
||||
from config.logger import setup_logging
|
||||
|
||||
TAG = __name__
|
||||
logger = setup_logging()
|
||||
|
||||
class IntentProviderBase(ABC):
|
||||
def __init__(self, config):
|
||||
self.config = config
|
||||
self.intent_options = config.get("intent_options", {
|
||||
"continue_chat": "继续聊天",
|
||||
"end_chat": "结束聊天",
|
||||
"play_music": "播放音乐"
|
||||
})
|
||||
|
||||
def set_llm(self, llm):
|
||||
self.llm = llm
|
||||
logger.bind(tag=TAG).debug("Set LLM for intent provider")
|
||||
|
||||
@abstractmethod
|
||||
async def detect_intent(self, dialogue_history: List[Dict]) -> str:
|
||||
"""
|
||||
检测用户最后一句话的意图
|
||||
Args:
|
||||
dialogue_history: 对话历史记录列表,每条记录包含role和content
|
||||
Returns:
|
||||
返回识别出的意图,格式为:
|
||||
- "继续聊天"
|
||||
- "结束聊天"
|
||||
- "播放音乐 歌名" 或 "随机播放音乐"
|
||||
"""
|
||||
pass
|
||||
@@ -0,0 +1,61 @@
|
||||
from typing import List, Dict
|
||||
from ..base import IntentProviderBase
|
||||
from config.logger import setup_logging
|
||||
|
||||
TAG = __name__
|
||||
logger = setup_logging()
|
||||
|
||||
|
||||
|
||||
class IntentProvider(IntentProviderBase):
|
||||
def __init__(self, config):
|
||||
super().__init__(config)
|
||||
self.llm = None
|
||||
self.promot = self.get_intent_system_prompt()
|
||||
|
||||
def get_intent_system_prompt(self) -> str:
|
||||
"""
|
||||
根据配置的意图选项动态生成系统提示词
|
||||
Returns:
|
||||
格式化后的系统提示词
|
||||
"""
|
||||
intent_list = []
|
||||
for key, value in self.intent_options.items():
|
||||
if key == "play_music":
|
||||
intent_list.append(f"{value} [歌名]")
|
||||
else:
|
||||
intent_list.append(value)
|
||||
|
||||
prompt = (
|
||||
"你是一个意图识别助手。你需要根据和用户的对话记录,重点分析用户的最后一句话,判断用户意图属于以下哪一类:\n"
|
||||
f"{', '.join(intent_list)}\n"
|
||||
"如果是唱歌、听歌、播放音乐,请指定歌名,格式为'播放音乐 [识别出的歌名]'。\n"
|
||||
"如果听不出具体歌名,可以返回'随机播放音乐'。\n"
|
||||
"只需要返回意图结果的json,不要解释。"
|
||||
"返回格式如下:\n"
|
||||
"{intent: '用户意图'}"
|
||||
)
|
||||
return prompt
|
||||
|
||||
async def detect_intent(self, dialogue_history: List[Dict]) -> str:
|
||||
if not self.llm:
|
||||
raise ValueError("LLM provider not set")
|
||||
|
||||
# 构建用户最后一句话的提示
|
||||
msgStr = ""
|
||||
for msg in dialogue_history:
|
||||
if msg.role == "user":
|
||||
msgStr += f"User: {msg.content}\n"
|
||||
elif msg.role== "assistant":
|
||||
msgStr += f"Assistant: {msg.content}\n"
|
||||
|
||||
user_prompt = f"请分析用户的意图:\n{msgStr}"
|
||||
|
||||
# 使用LLM进行意图识别
|
||||
intent = self.llm.response_no_stream(
|
||||
system_prompt=self.promot,
|
||||
user_prompt=user_prompt
|
||||
)
|
||||
|
||||
logger.bind(tag=TAG).info(f"Detected intent: {intent}")
|
||||
return intent.strip()
|
||||
@@ -0,0 +1,18 @@
|
||||
from ..base import IntentProviderBase
|
||||
from typing import List, Dict
|
||||
from config.logger import setup_logging
|
||||
|
||||
TAG = __name__
|
||||
logger = setup_logging()
|
||||
|
||||
class IntentProvider(IntentProviderBase):
|
||||
async def detect_intent(self, dialogue_history: List[Dict]) -> str:
|
||||
"""
|
||||
默认的意图识别实现,始终返回继续聊天
|
||||
Args:
|
||||
dialogue_history: 对话历史记录列表
|
||||
Returns:
|
||||
固定返回"继续聊天"
|
||||
"""
|
||||
logger.bind(tag=TAG).debug("Using NoIntentProvider, always returning continue chat")
|
||||
return self.intent_options["continue_chat"]
|
||||
Reference in New Issue
Block a user