Hass/xiaozhi-esp32-server
1
0
Fork 0
mirror of https://github.com/xinnan-tech/xiaozhi-esp32-server.git synced 2026-07-28 10:03:54 +08:00
Code Issues Packages Projects Releases Wiki Activity

merge main

Browse Source
This commit is contained in:
lizhongxiang
2025-03-13 14:24:18 +08:00
parent 0e8edfe002 a8f06f5718
commit 91bdc05a38
52 changed files with 1013 additions and 1184 deletions
Download Patch File Download Diff File Expand all files Collapse all files
main/xiaozhi-server/core/providers/intent/base.py
+3 -2
View File
@@ -5,6 +5,7 @@ from config.logger import setup_logging
TAG = __name__
logger = setup_logging()
class IntentProviderBase(ABC):
def __init__(self, config):
self.config = config
@@ -17,9 +18,9 @@ class IntentProviderBase(ABC):
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:
async def detect_intent(self, conn, dialogue_history: List[Dict], text: str) -> str:
"""
检测用户最后一句话的意图
Args:
main/xiaozhi-server/core/providers/intent/intent_llm/intent_llm.py
+69 -21
View File
@@ -1,7 +1,7 @@
from typing import List, Dict
from ..base import IntentProviderBase
from config.logger import setup_logging
import re
TAG = __name__
logger = setup_logging()
@@ -20,42 +20,90 @@ class IntentProvider(IntentProviderBase):
格式化后的系统提示词
"""
intent_list = []
"""
"continue_chat": "1.继续聊天, 除了播放音乐和结束聊天的时候的选项, 比如日常的聊天和问候, 对话等",
"end_chat": "2.结束聊天, 用户发来如再见之类的表示结束的话, 不想再进行对话的时候",
"play_music": "3.播放音乐, 用户希望你可以播放音乐, 只用于播放音乐的意图"
"""
for key, value in self.intent_options.items():
if key == "play_music":
intent_list.append(f"{value} [歌名]")
intent_list.append("3.播放音乐, 用户希望你可以播放音乐, 只用于播放音乐的意图")
elif key == "end_chat":
intent_list.append("2.结束聊天, 用户发来如再见之类的表示结束的话, 不想再进行对话的时候")
elif key == "continue_chat":
intent_list.append("1.继续聊天, 除了播放音乐和结束聊天的时候的选项, 比如日常的聊天和问候, 对话等")
else:
intent_list.append(value)
# "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"
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'}"
"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(使用<start>和<end>标志)：\n"
"<start>"
f"{', '.join(intent_list)}"
"<end>\n"
"你需要按照以下的步骤处理用户的对话"
"1. 思考出对话的意图是哪一类的"
"2. 属于1和2的意图, 直接返回，返回格式如下：\n"
"{intent: '用户意图'}\n"
"3. 属于3的意图，则继续分析用户希望播放的音乐\n"
"4. 如果无法识别出具体歌名，可以返回'随机播放音乐'\n"
"{intent: '播放音乐 [获取的音乐名字]'}\n"
"下面是几个处理的示例(思考的内容不返回, 只返回json部分, 无额外的内容)\n"
"```"
"用户: 你今天怎么样?\n"
"思考(不返回): 用户发来的数据是一个问候语，属于继续聊天的意图, 是种类1, 种类1的需求是直接返回\n"
"返回结果: {intent: '继续聊天'}\n"
"```"
"用户: 我今天有点累了, 我们明天再聊吧\n"
"思考(不返回): 用户表达了今天不想继续对话，属于结束聊天的意图, 是种类2, 种类2的需求是直接返回\n"
"返回结果: {intent: '结束聊天'}\n"
"```"
"用户: 我今天有点累了, 我们明天再聊吧\n"
"思考(不返回): 用户表达了今天不想继续对话，属于结束聊天的意图, 是种类2, 种类2的需求是直接返回\n"
"返回结果: {intent: '结束聊天'}\n"
"```"
"用户: 你可以播放一首中秋月给我听吗\n"
"思考(不返回): 用户表达了想听音乐的续签，属于播放音乐的意图, 是种类3, 种类3的需求需要继续判断播放的音乐, 这里用户希望的歌曲名明确给出是中秋月\n"
"返回结果: {intent: '播放音乐 [中秋月]'}\n"
"```"
"你现在可以使用的音乐的名称如下(使用<start>和<end>标志):\n"
)
return prompt
async def detect_intent(self, dialogue_history: List[Dict]) -> str:
async def detect_intent(self, conn, dialogue_history: List[Dict], text:str) -> 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}"
# 只使用最后两句即可
if len(dialogue_history) >= 2:
# 保证最少有两句话的时候处理
msgStr += f"{dialogue_history[-2].role}: {dialogue_history[-2].content}\n"
msgStr += f"{dialogue_history[-1].role}: {dialogue_history[-1].content}\n"
msgStr += f"User: {text}\n"
user_prompt = f"当前的对话如下：\n{msgStr}"
prompt_music = f"{self.promot}\n<start>{conn.music_handler.music_files}\n<end>"
logger.bind(tag=TAG).debug(f"User prompt: {prompt_music}")
# 使用LLM进行意图识别
intent = self.llm.response_no_stream(
system_prompt=self.promot,
system_prompt=prompt_music,
user_prompt=user_prompt
)
# 使用正则表达式提取大括号中的内容
# 使用正则表达式提取 {} 中的内容
match = re.search(r'\{.*?\}', intent)
if match:
result = match.group(0) # 获取匹配到的内容（包含 {}）
print(result) # 输出：{intent: '播放音乐 [中秋月]'}
intent = result
else:
intent = "{intent: '继续聊天'}"
logger.bind(tag=TAG).info(f"Detected intent: {intent}")
return intent.strip()
return intent.strip()
main/xiaozhi-server/core/providers/intent/nointent/nointent.py
+3 -1
View File
@@ -5,12 +5,14 @@ from config.logger import setup_logging
TAG = __name__
logger = setup_logging()
class IntentProvider(IntentProviderBase):
async def detect_intent(self, dialogue_history: List[Dict]) -> str:
async def detect_intent(self, conn, dialogue_history: List[Dict], text: str) -> 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
text: 本次对话记录
Returns:
固定返回"继续聊天"
"""
main/xiaozhi-server/core/providers/llm/coze/coze.py
+14 -1
View File
@@ -11,19 +11,31 @@ from cozepy import Coze, TokenAuth, Message, ChatStatus, MessageContentType, Cha
TAG = __name__
logger = setup_logging()
class LLMProvider(LLMProviderBase):
def __init__(self, config):
self.personal_access_token = config.get("personal_access_token")
self.bot_id = config.get("bot_id")
self.user_id = config.get("user_id")
self.session_conversation_map = {} # 存储session_id和conversation_id的映射
def response(self, session_id, dialogue):
coze_api_token = self.personal_access_token
coze_api_base = COZE_CN_BASE_URL
last_msg = next(m for m in reversed(dialogue) if m["role"] == "user")
coze = Coze(auth=TokenAuth(token=coze_api_token), base_url=coze_api_base)
conversation_id = self.session_conversation_map.get(session_id)
# 如果没有找到conversation_id，则创建新的对话
if not conversation_id:
conversation = coze.conversations.create(
messages=[
]
)
conversation_id = conversation.id
self.session_conversation_map[session_id] = conversation_id # 更新映射
for event in coze.chat.stream(
bot_id=self.bot_id,
@@ -31,6 +43,7 @@ class LLMProvider(LLMProviderBase):
additional_messages=[
Message.build_user_question_text(last_msg["content"]),
],
conversation_id=conversation_id,
):
if event.event == ChatEventType.CONVERSATION_MESSAGE_DELTA:
print(event.message.content, end="", flush=True)
main/xiaozhi-server/core/providers/llm/ollama/ollama.py
+1 -23
View File
@@ -51,30 +51,8 @@ class LLMProvider(LLMProviderBase):
tools=functions,
)
current_function_call = None
current_content = ""
for chunk in stream:
delta = chunk.choices[0].delta
if delta.content:
current_content += delta.content
yield {"type": "content", "content": delta.content}
if delta.tool_calls:
tool_call = delta.tool_calls[0]
# Handle the function call data using proper attribute access
if not current_function_call:
current_function_call = {
"function": {
"name": tool_call.function.name,
"arguments": tool_call.function.arguments
}
}
if current_function_call:
logger.bind(tag=TAG).debug(f"ollama Function call detected: {current_function_call}")
yield {"type": "function_call", "function_call": current_function_call}
yield chunk.choices[0].delta.content, chunk.choices[0].delta.tool_calls
except Exception as e:
logger.bind(tag=TAG).error(f"Error in Ollama function call: {e}")
main/xiaozhi-server/core/providers/llm/openai/openai.py
+1 -23
View File
@@ -53,30 +53,8 @@ class LLMProvider(LLMProviderBase):
tools=functions,
)
current_function_call = None
current_content = ""
for chunk in stream:
delta = chunk.choices[0].delta
if delta.content:
current_content += delta.content
yield {"type": "content", "content": delta.content}
if delta.tool_calls:
tool_call = delta.tool_calls[0]
# Handle the function call data using proper attribute access
if not current_function_call:
current_function_call = {
"function": {
"name": tool_call.function.name,
"arguments": tool_call.function.arguments
}
}
if current_function_call:
logger.bind(tag=TAG).debug(f"openai Function call detected: {current_function_call}")
yield {"type": "function_call", "function_call": current_function_call}
yield chunk.choices[0].delta.content, chunk.choices[0].delta.tool_calls
except Exception as e:
self.logger.bind(tag=TAG).error(f"Error in function call streaming: {e}")