mirror of
https://github.com/xinnan-tech/xiaozhi-esp32-server.git
synced 2026-07-26 00:53:54 +08:00
Merge pull request #3155 from xinnan-tech/perf-tool-call-optimization
perf：优化工具调用问题、调整后端智能体搜索条件逻辑
This commit is contained in:
wengzh
@@ -52,6 +52,26 @@ auto_import_modules("plugins_func.functions")
|
||||
class TTSException(RuntimeError):
|
||||
pass
|
||||
|
||||
# direct_answer 虚拟工具定义
|
||||
# 不是真实工具，是路由机制：将"调不调工具"的二选一变为"调哪个"的多选，防止小模型误触发真实工具
|
||||
DIRECT_ANSWER_TOOL = {
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "direct_answer",
|
||||
"description": "当用户的请求不匹配其他任何工具时，可用此选项直接回复。将回复内容写在response参数里。",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"response": {
|
||||
"type": "string",
|
||||
"description": "你回复用户的完整内容",
|
||||
},
|
||||
},
|
||||
"required": ["response"],
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
class ConnectionHandler:
|
||||
def __init__(
|
||||
@@ -520,9 +540,10 @@ class ConnectionHandler:
|
||||
self.logger.bind(tag=TAG).debug("系统提示词已增强更新")
|
||||
|
||||
def _inject_tool_call_fewshot(self):
|
||||
"""注入工具调用 few-shot 示例到对话历史，帮助模型建立正确的工具调用三段式模式。
|
||||
参考 Action.RECORD 的 assistant(tool_calls) → tool(result) → assistant(response) 模式，
|
||||
让模型从第一轮对话就能看到正确的工具调用行为，避免上下文污染导致工具调用退化。
|
||||
"""注入工具调用 few-shot 示例到对话历史。
|
||||
结构：正样本（工具调用示例）放在动态 system 之前，可命中前缀缓存；
|
||||
负样本（直接回答示例）放在动态 system 之后、紧挨真实用户消息，
|
||||
确保模型在处理用户消息前最后看到的是"不调工具"的行为模式。
|
||||
"""
|
||||
if self.intent_type != "function_call":
|
||||
return
|
||||
@@ -533,9 +554,29 @@ class ConnectionHandler:
|
||||
if not tools:
|
||||
return
|
||||
|
||||
# 根据可用工具动态构建 few-shot 示例
|
||||
tool_names = {t.get("function", {}).get("name") for t in tools}
|
||||
|
||||
# === few-shot 示例（is_temporary）===
|
||||
# 展示 direct_answer 携带 response 参数的用法，一次调用完成回复
|
||||
|
||||
# 示例1：direct_answer（回复内容写在 response 参数里，无需递归）
|
||||
da_tc_id = "fewshot_da_001"
|
||||
self.dialogue.put(Message(role="user", content="给我讲个故事吧", is_temporary=True))
|
||||
self.dialogue.put(Message(
|
||||
role="assistant",
|
||||
tool_calls=[{
|
||||
"id": da_tc_id,
|
||||
"function": {"arguments": '{"response": "好呀，你想听什么类型的呀？童话、冒险还是搞笑的？选一个我给你开讲~"}', "name": "direct_answer"},
|
||||
"type": "function", "index": 0,
|
||||
}],
|
||||
is_temporary=True,
|
||||
))
|
||||
self.dialogue.put(Message(
|
||||
role="tool", tool_call_id=da_tc_id,
|
||||
content="已直接回复", is_temporary=True,
|
||||
))
|
||||
|
||||
# 示例2：真实工具调用（handle_exit_intent）
|
||||
if "handle_exit_intent" in tool_names:
|
||||
tc_id = "fewshot_exit_001"
|
||||
self.dialogue.put(Message(role="user", content="拜拜", is_temporary=True))
|
||||
@@ -556,43 +597,6 @@ class ConnectionHandler:
|
||||
role="assistant", content="再见，下次再聊~", is_temporary=True,
|
||||
))
|
||||
|
||||
if "play_music" in tool_names:
|
||||
tc_id = "fewshot_music_001"
|
||||
self.dialogue.put(Message(role="user", content="放首歌", is_temporary=True))
|
||||
self.dialogue.put(Message(
|
||||
role="assistant",
|
||||
tool_calls=[{
|
||||
"id": tc_id,
|
||||
"function": {"arguments": '{"song_name": "random"}', "name": "play_music"},
|
||||
"type": "function", "index": 0,
|
||||
}],
|
||||
is_temporary=True,
|
||||
))
|
||||
self.dialogue.put(Message(
|
||||
role="tool", tool_call_id=tc_id,
|
||||
content="正在为您播放音乐", is_temporary=True,
|
||||
))
|
||||
self.dialogue.put(Message(
|
||||
role="assistant", content="好嘞，给你安排上~", is_temporary=True,
|
||||
))
|
||||
|
||||
# 负向示例：用户请求普通对话内容时，直接回答，不调用任何工具
|
||||
# 帮助弱模型建立"该调才调、不该调不调"的判断能力
|
||||
# 注意：示例回复不能包含具体的创作内容（故事、诗歌等），
|
||||
# 否则弱模型会直接复述示例内容，而无法泛化出正确的行为模式
|
||||
self.dialogue.put(Message(role="user", content="给我讲个故事吧", is_temporary=True))
|
||||
self.dialogue.put(Message(
|
||||
role="assistant",
|
||||
content="好呀，你想听什么类型的呀？童话、冒险还是搞笑的？选一个我给你开讲~",
|
||||
is_temporary=True,
|
||||
))
|
||||
self.dialogue.put(Message(role="user", content="你知道为什么天空是蓝色的吗", is_temporary=True))
|
||||
self.dialogue.put(Message(
|
||||
role="assistant",
|
||||
content="天空看起来是蓝色，是因为阳光穿过大气层的时候，蓝色光波长短，被空气分子散射得最厉害，所以我们抬头一看就是满眼蓝色啦。",
|
||||
is_temporary=True,
|
||||
))
|
||||
|
||||
self.logger.bind(tag=TAG).debug("已注入工具调用 few-shot 示例")
|
||||
|
||||
def _init_report_threads(self):
|
||||
@@ -948,6 +952,10 @@ class ConnectionHandler:
|
||||
and not force_final_answer
|
||||
):
|
||||
functions = self.func_handler.get_functions()
|
||||
# 仅在第一层调用时注入 direct_answer 虚拟工具
|
||||
# 递归调用（depth>0）不注入，避免模型在生成文本回复时再次调 direct_answer 导致循环
|
||||
if functions is not None and depth == 0:
|
||||
functions.append(DIRECT_ANSWER_TOOL)
|
||||
|
||||
response_message = []
|
||||
|
||||
@@ -1006,6 +1014,23 @@ class ConnectionHandler:
|
||||
if tools_call is not None and len(tools_call) > 0:
|
||||
tool_call_flag = True
|
||||
self._merge_tool_calls(tool_calls_list, tools_call)
|
||||
|
||||
# 流式提取 direct_answer 的 response 参数，实时送 TTS
|
||||
for tc in tool_calls_list:
|
||||
if tc["name"] == "direct_answer" and tc.get("arguments"):
|
||||
da_text = self._extract_direct_answer_response(tc["arguments"])
|
||||
sent_len = tc.get("_da_sent", 0)
|
||||
if da_text and len(da_text) > sent_len:
|
||||
new_part = da_text[sent_len:]
|
||||
tc["_da_sent"] = len(da_text)
|
||||
self.tts.tts_text_queue.put(
|
||||
TTSMessageDTO(
|
||||
sentence_id=current_sentence_id,
|
||||
sentence_type=SentenceType.MIDDLE,
|
||||
content_type=ContentType.TEXT,
|
||||
content_detail=new_part,
|
||||
)
|
||||
)
|
||||
else:
|
||||
content = response
|
||||
|
||||
@@ -1077,6 +1102,35 @@ class ConnectionHandler:
|
||||
f"function call error: {content_arguments}"
|
||||
)
|
||||
|
||||
if not bHasError and len(tool_calls_list) > 0:
|
||||
# 处理 direct_answer 虚拟工具
|
||||
direct_answer_calls = [tc for tc in tool_calls_list if tc["name"] == "direct_answer"]
|
||||
real_tool_calls = [tc for tc in tool_calls_list if tc["name"] != "direct_answer"]
|
||||
|
||||
if direct_answer_calls:
|
||||
self.logger.bind(tag=TAG).debug(
|
||||
f"模型选择 direct_answer，流式已播报，写入对话历史"
|
||||
)
|
||||
for tc in direct_answer_calls:
|
||||
# 文本已在流式阶段实时播报，这里只写入对话历史
|
||||
da_response = self._extract_direct_answer_response(tc.get("arguments", "{}"))
|
||||
if da_response:
|
||||
self.tts.store_tts_text(current_sentence_id, da_response)
|
||||
self.dialogue.put(Message(role="assistant", content=da_response))
|
||||
|
||||
if not real_tool_calls:
|
||||
if depth == 0:
|
||||
self.tts.tts_text_queue.put(
|
||||
TTSMessageDTO(
|
||||
sentence_id=current_sentence_id,
|
||||
sentence_type=SentenceType.LAST,
|
||||
content_type=ContentType.ACTION,
|
||||
)
|
||||
)
|
||||
return
|
||||
|
||||
tool_calls_list = real_tool_calls
|
||||
|
||||
if not bHasError and len(tool_calls_list) > 0:
|
||||
self.logger.bind(tag=TAG).debug(
|
||||
f"检测到 {len(tool_calls_list)} 个工具调用"
|
||||
@@ -1493,6 +1547,33 @@ class ConnectionHandler:
|
||||
finally:
|
||||
self.logger.bind(tag=TAG).info("超时检查任务已退出")
|
||||
|
||||
@staticmethod
|
||||
def _extract_direct_answer_response(arguments_str):
|
||||
"""从 direct_answer 的参数中提取 response 值。
|
||||
支持完整的 JSON 和流式中不完整的 JSON 片段。
|
||||
"""
|
||||
if not arguments_str:
|
||||
return ""
|
||||
# 找到 "response": " 的位置
|
||||
marker = '"response": "'
|
||||
idx = arguments_str.find(marker)
|
||||
if idx < 0:
|
||||
# 尝试不带空格的格式
|
||||
marker = '"response":"'
|
||||
idx = arguments_str.find(marker)
|
||||
if idx < 0:
|
||||
return ""
|
||||
start = idx + len(marker)
|
||||
raw = arguments_str[start:]
|
||||
# 去掉末尾的 JSON 闭合符号（如果已完整）
|
||||
if raw.endswith('"}'):
|
||||
raw = raw[:-2]
|
||||
elif raw.endswith('"'):
|
||||
raw = raw[:-1]
|
||||
# 处理 JSON 转义
|
||||
raw = raw.replace('\\"', '"').replace('\\n', '\n').replace('\\\\', '\\')
|
||||
return raw
|
||||
|
||||
def _merge_tool_calls(self, tool_calls_list, tools_call):
|
||||
"""合并工具调用列表
|
||||
|
||||
|
||||
Reference in New Issue
Block a user