mirror of https://github.com/Mai-with-u/MaiBot.git
feat；优化了记忆检索的速度和token消耗（将question提出交给planenr)
SengokuCola
parent
490589b0ad
commit
0852af49f9
|
|
@ -1,5 +1,9 @@
|
|||
# Changelog
|
||||
|
||||
移除 enable_jargon_detection
|
||||
添加 global_memory_blacklist
|
||||
|
||||
|
||||
## [0.12.0] - 2025-12-21
|
||||
### 🌟 重大更新
|
||||
- 添加思考力度机制，动态控制回复时间和长度
|
||||
|
|
|
|||
|
|
@ -407,8 +407,8 @@ class ExpressionSelector:
|
|||
# 4. 调用LLM
|
||||
content, (reasoning_content, model_name, _) = await self.llm_model.generate_response_async(prompt=prompt)
|
||||
|
||||
print(prompt)
|
||||
print(content)
|
||||
# print(prompt)
|
||||
# print(content)
|
||||
|
||||
if not content:
|
||||
logger.warning("LLM返回空结果")
|
||||
|
|
|
|||
|
|
@ -341,7 +341,7 @@ async def retrieve_concepts_with_jargon(concepts: List[str], chat_id: str) -> st
|
|||
meaning = result.get("meaning", "").strip()
|
||||
if found_content and meaning:
|
||||
output_parts.append(f"找到 '{found_content}' 的含义为：{meaning}")
|
||||
results.append("，".join(output_parts))
|
||||
results.append("\n".join(output_parts)) # 换行分隔每个jargon解释
|
||||
logger.info(f"在jargon库中找到匹配（模糊搜索）: {concept}，找到{len(jargon_results)}条结果")
|
||||
else:
|
||||
# 精确匹配
|
||||
|
|
@ -350,7 +350,8 @@ async def retrieve_concepts_with_jargon(concepts: List[str], chat_id: str) -> st
|
|||
meaning = result.get("meaning", "").strip()
|
||||
if meaning:
|
||||
output_parts.append(f"'{concept}' 为黑话或者网络简写，含义为：{meaning}")
|
||||
results.append("；".join(output_parts) if len(output_parts) > 1 else output_parts[0])
|
||||
# 换行分隔每个jargon解释
|
||||
results.append("\n".join(output_parts) if len(output_parts) > 1 else output_parts[0])
|
||||
exact_matches.append(concept) # 收集精确匹配的概念，稍后统一打印
|
||||
else:
|
||||
# 未找到，不返回占位信息，只记录日志
|
||||
|
|
@ -361,5 +362,5 @@ async def retrieve_concepts_with_jargon(concepts: List[str], chat_id: str) -> st
|
|||
logger.info(f"找到黑话: {', '.join(exact_matches)}，共找到{len(exact_matches)}条结果")
|
||||
|
||||
if results:
|
||||
return "【概念检索结果】\n" + "\n".join(results) + "\n"
|
||||
return "你了解以下词语可能的含义：\n" + "\n".join(results) + "\n"
|
||||
return ""
|
||||
|
|
|
|||
|
|
@ -687,7 +687,7 @@ class HeartFChatting:
|
|||
return {
|
||||
"action_type": "reply",
|
||||
"success": True,
|
||||
"result": f"你回复内容{reply_text}",
|
||||
"result": f"你使用reply动作，对' {action_planner_info.action_message.processed_plain_text} '这句话进行了回复，回复内容为: '{reply_text}'",
|
||||
"loop_info": loop_info,
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -53,7 +53,7 @@ reply
|
|||
4.不要选择回复你自己发送的消息
|
||||
5.不要单独对表情包进行回复
|
||||
6.将上下文中所有含义不明的，疑似黑话的，缩写词均写入unknown_words中
|
||||
7.用一句简单的话来描述当前回复场景，不超过10个字
|
||||
7.如果你对上下文存在疑问，有需要查询的问题，写入question中
|
||||
{reply_action_example}
|
||||
|
||||
no_reply
|
||||
|
|
@ -224,6 +224,25 @@ class ActionPlanner:
|
|||
else:
|
||||
reasoning = "未提供原因"
|
||||
action_data = {key: value for key, value in action_json.items() if key not in ["action"]}
|
||||
|
||||
# 验证和清理 question
|
||||
if "question" in action_data:
|
||||
q = action_data.get("question")
|
||||
if isinstance(q, str):
|
||||
cleaned_q = q.strip()
|
||||
if cleaned_q:
|
||||
action_data["question"] = cleaned_q
|
||||
else:
|
||||
# 如果清理后为空字符串，移除该字段
|
||||
action_data.pop("question", None)
|
||||
elif q is None:
|
||||
# 如果为 None，移除该字段
|
||||
action_data.pop("question", None)
|
||||
else:
|
||||
# 如果不是字符串类型，记录警告并移除
|
||||
logger.warning(f"{self.log_prefix}question 格式不正确，应为字符串类型，已忽略")
|
||||
action_data.pop("question", None)
|
||||
|
||||
# 非no_reply动作需要target_message_id
|
||||
target_message = None
|
||||
|
||||
|
|
@ -503,18 +522,20 @@ class ActionPlanner:
|
|||
name_block = f"你的名字是{bot_name}{bot_nickname}，请注意哪些是你自己的发言。"
|
||||
|
||||
# 根据 think_mode 配置决定 reply action 的示例 JSON
|
||||
# 在 JSON 中直接作为 action 参数携带 unknown_words
|
||||
# 在 JSON 中直接作为 action 参数携带 unknown_words 和 question
|
||||
if global_config.chat.think_mode == "classic":
|
||||
reply_action_example = (
|
||||
'{{"action":"reply", "target_message_id":"消息id(m+数字)", '
|
||||
'"unknown_words":["词语1","词语2"]}}'
|
||||
'"unknown_words":["词语1","词语2"], '
|
||||
'"question":"需要查询的问题"}'
|
||||
)
|
||||
else:
|
||||
reply_action_example = (
|
||||
"5.think_level表示思考深度，0表示该回复不需要思考和回忆，1表示该回复需要进行回忆和思考\n"
|
||||
+ '{{"action":"reply", "think_level":数值等级(0或1), '
|
||||
'"target_message_id":"消息id(m+数字)", '
|
||||
'"unknown_words":["词语1","词语2"]}}'
|
||||
'"unknown_words":["词语1","词语2"], '
|
||||
'"question":"需要查询的问题"}'
|
||||
)
|
||||
|
||||
planner_prompt_template = await global_prompt_manager.get_prompt_async("planner_prompt")
|
||||
|
|
|
|||
|
|
@ -947,6 +947,18 @@ class DefaultReplyer:
|
|||
chat_id, message_list_before_short, chat_talking_prompt_short, unknown_words
|
||||
)
|
||||
|
||||
# 从 chosen_actions 中提取 question（仅在 reply 动作中）
|
||||
question = None
|
||||
if chosen_actions:
|
||||
for action_info in chosen_actions:
|
||||
if action_info.action_type == "reply" and isinstance(action_info.action_data, dict):
|
||||
q = action_info.action_data.get("question")
|
||||
if isinstance(q, str):
|
||||
cleaned_q = q.strip()
|
||||
if cleaned_q:
|
||||
question = cleaned_q
|
||||
break
|
||||
|
||||
# 并行执行构建任务（包括黑话解释，可配置关闭）
|
||||
task_results = await asyncio.gather(
|
||||
self._time_and_run_task(
|
||||
|
|
@ -961,7 +973,7 @@ class DefaultReplyer:
|
|||
self._time_and_run_task(self.build_personality_prompt(), "personality_prompt"),
|
||||
self._time_and_run_task(
|
||||
build_memory_retrieval_prompt(
|
||||
chat_talking_prompt_short, sender, target, self.chat_stream, think_level=think_level
|
||||
chat_talking_prompt_short, sender, target, self.chat_stream, think_level=think_level, unknown_words=unknown_words, question=question
|
||||
),
|
||||
"memory_retrieval",
|
||||
),
|
||||
|
|
|
|||
|
|
@ -110,6 +110,7 @@ class PrivateReplyer:
|
|||
enable_tool=enable_tool,
|
||||
reply_message=reply_message,
|
||||
reply_reason=reply_reason,
|
||||
unknown_words=unknown_words,
|
||||
)
|
||||
llm_response.prompt = prompt
|
||||
llm_response.selected_expressions = selected_expressions
|
||||
|
|
@ -611,6 +612,7 @@ class PrivateReplyer:
|
|||
available_actions: Optional[Dict[str, ActionInfo]] = None,
|
||||
chosen_actions: Optional[List[ActionPlannerInfo]] = None,
|
||||
enable_tool: bool = True,
|
||||
unknown_words: Optional[List[str]] = None,
|
||||
) -> Tuple[str, List[int]]:
|
||||
"""
|
||||
构建回复器上下文
|
||||
|
|
@ -709,12 +711,24 @@ class PrivateReplyer:
|
|||
else:
|
||||
jargon_coroutine = self._build_disabled_jargon_explanation()
|
||||
|
||||
# 从 chosen_actions 中提取 question（仅在 reply 动作中）
|
||||
question = None
|
||||
if chosen_actions:
|
||||
for action_info in chosen_actions:
|
||||
if action_info.action_type == "reply" and isinstance(action_info.action_data, dict):
|
||||
q = action_info.action_data.get("question")
|
||||
if isinstance(q, str):
|
||||
cleaned_q = q.strip()
|
||||
if cleaned_q:
|
||||
question = cleaned_q
|
||||
break
|
||||
|
||||
# 并行执行九个构建任务（包括黑话解释，可配置关闭）
|
||||
task_results = await asyncio.gather(
|
||||
self._time_and_run_task(
|
||||
self.build_expression_habits(chat_talking_prompt_short, target, reply_reason), "expression_habits"
|
||||
),
|
||||
self._time_and_run_task(self.build_relation_info(chat_talking_prompt_short, sender), "relation_info"),
|
||||
# self._time_and_run_task(self.build_relation_info(chat_talking_prompt_short, sender), "relation_info"),
|
||||
self._time_and_run_task(
|
||||
self.build_tool_info(chat_talking_prompt_short, sender, target, enable_tool=enable_tool), "tool_info"
|
||||
),
|
||||
|
|
@ -723,7 +737,7 @@ class PrivateReplyer:
|
|||
self._time_and_run_task(self.build_personality_prompt(), "personality_prompt"),
|
||||
self._time_and_run_task(
|
||||
build_memory_retrieval_prompt(
|
||||
chat_talking_prompt_short, sender, target, self.chat_stream, self.tool_executor
|
||||
chat_talking_prompt_short, sender, target, self.chat_stream, think_level=1, unknown_words=unknown_words, question=question
|
||||
),
|
||||
"memory_retrieval",
|
||||
),
|
||||
|
|
|
|||
|
|
@ -743,13 +743,13 @@ class StatisticOutputTask(AsyncTask):
|
|||
"""
|
||||
if stats[TOTAL_REQ_CNT] <= 0:
|
||||
return ""
|
||||
data_fmt = "{:<32} {:>10} {:>12} {:>12} {:>12} {:>9.2f}¥ {:>10.1f} {:>10.1f} {:>12} {:>12}"
|
||||
data_fmt = "{:<32} {:>10} {:>12} {:>12} {:>12} {:>9.2f}¥ {:>10.1f} {:>10.1f} {:>12} {:>12} {:>12}"
|
||||
|
||||
total_replies = stats.get(TOTAL_REPLY_CNT, 0)
|
||||
|
||||
output = [
|
||||
"按模型分类统计:",
|
||||
" 模型名称 调用次数 输入Token 输出Token Token总量 累计花费 平均耗时(秒) 标准差(秒) 每次回复平均调用次数 每次回复平均Token数",
|
||||
" 模型名称 调用次数 输入Token 输出Token Token总量 累计花费 平均耗时(秒) 标准差(秒) 每次回复平均调用次数 每次回复平均Token数 每次调用平均Token",
|
||||
]
|
||||
for model_name, count in sorted(stats[REQ_CNT_BY_MODEL].items()):
|
||||
name = f"{model_name[:29]}..." if len(model_name) > 32 else model_name
|
||||
|
|
@ -764,6 +764,9 @@ class StatisticOutputTask(AsyncTask):
|
|||
avg_count_per_reply = count / total_replies if total_replies > 0 else 0.0
|
||||
avg_tokens_per_reply = tokens / total_replies if total_replies > 0 else 0.0
|
||||
|
||||
# 计算每次调用平均token
|
||||
avg_tokens_per_call = tokens / count if count > 0 else 0.0
|
||||
|
||||
# 格式化大数字
|
||||
formatted_count = _format_large_number(count)
|
||||
formatted_in_tokens = _format_large_number(in_tokens)
|
||||
|
|
@ -771,6 +774,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
formatted_tokens = _format_large_number(tokens)
|
||||
formatted_avg_count = _format_large_number(avg_count_per_reply) if total_replies > 0 else "N/A"
|
||||
formatted_avg_tokens = _format_large_number(avg_tokens_per_reply) if total_replies > 0 else "N/A"
|
||||
formatted_avg_tokens_per_call = _format_large_number(avg_tokens_per_call) if count > 0 else "N/A"
|
||||
|
||||
output.append(
|
||||
data_fmt.format(
|
||||
|
|
@ -784,6 +788,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
std_time_cost,
|
||||
formatted_avg_count,
|
||||
formatted_avg_tokens,
|
||||
formatted_avg_tokens_per_call,
|
||||
)
|
||||
)
|
||||
|
||||
|
|
@ -797,13 +802,13 @@ class StatisticOutputTask(AsyncTask):
|
|||
"""
|
||||
if stats[TOTAL_REQ_CNT] <= 0:
|
||||
return ""
|
||||
data_fmt = "{:<32} {:>10} {:>12} {:>12} {:>12} {:>9.2f}¥ {:>10.1f} {:>10.1f} {:>12} {:>12}"
|
||||
data_fmt = "{:<32} {:>10} {:>12} {:>12} {:>12} {:>9.2f}¥ {:>10.1f} {:>10.1f} {:>12} {:>12} {:>12}"
|
||||
|
||||
total_replies = stats.get(TOTAL_REPLY_CNT, 0)
|
||||
|
||||
output = [
|
||||
"按模块分类统计:",
|
||||
" 模块名称 调用次数 输入Token 输出Token Token总量 累计花费 平均耗时(秒) 标准差(秒) 每次回复平均调用次数 每次回复平均Token数",
|
||||
" 模块名称 调用次数 输入Token 输出Token Token总量 累计花费 平均耗时(秒) 标准差(秒) 每次回复平均调用次数 每次回复平均Token数 每次调用平均Token",
|
||||
]
|
||||
for module_name, count in sorted(stats[REQ_CNT_BY_MODULE].items()):
|
||||
name = f"{module_name[:29]}..." if len(module_name) > 32 else module_name
|
||||
|
|
@ -818,6 +823,9 @@ class StatisticOutputTask(AsyncTask):
|
|||
avg_count_per_reply = count / total_replies if total_replies > 0 else 0.0
|
||||
avg_tokens_per_reply = tokens / total_replies if total_replies > 0 else 0.0
|
||||
|
||||
# 计算每次调用平均token
|
||||
avg_tokens_per_call = tokens / count if count > 0 else 0.0
|
||||
|
||||
# 格式化大数字
|
||||
formatted_count = _format_large_number(count)
|
||||
formatted_in_tokens = _format_large_number(in_tokens)
|
||||
|
|
@ -825,6 +833,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
formatted_tokens = _format_large_number(tokens)
|
||||
formatted_avg_count = _format_large_number(avg_count_per_reply) if total_replies > 0 else "N/A"
|
||||
formatted_avg_tokens = _format_large_number(avg_tokens_per_reply) if total_replies > 0 else "N/A"
|
||||
formatted_avg_tokens_per_call = _format_large_number(avg_tokens_per_call) if count > 0 else "N/A"
|
||||
|
||||
output.append(
|
||||
data_fmt.format(
|
||||
|
|
@ -838,6 +847,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
std_time_cost,
|
||||
formatted_avg_count,
|
||||
formatted_avg_tokens,
|
||||
formatted_avg_tokens_per_call,
|
||||
)
|
||||
)
|
||||
|
||||
|
|
@ -935,11 +945,12 @@ class StatisticOutputTask(AsyncTask):
|
|||
f"<td>{stat_data[STD_TIME_COST_BY_MODEL][model_name]:.1f} 秒</td>"
|
||||
f"<td>{_format_large_number(count / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_MODEL][model_name] / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_MODEL][model_name] / count, html=True) if count > 0 else 'N/A'}</td>"
|
||||
f"</tr>"
|
||||
for model_name, count in sorted(stat_data[REQ_CNT_BY_MODEL].items())
|
||||
]
|
||||
if stat_data[REQ_CNT_BY_MODEL]
|
||||
else ["<tr><td colspan='10' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
else ["<tr><td colspan='11' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
)
|
||||
# 按请求类型分类统计
|
||||
type_rows = "\n".join(
|
||||
|
|
@ -955,11 +966,12 @@ class StatisticOutputTask(AsyncTask):
|
|||
f"<td>{stat_data[STD_TIME_COST_BY_TYPE][req_type]:.1f} 秒</td>"
|
||||
f"<td>{_format_large_number(count / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_TYPE][req_type] / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_TYPE][req_type] / count, html=True) if count > 0 else 'N/A'}</td>"
|
||||
f"</tr>"
|
||||
for req_type, count in sorted(stat_data[REQ_CNT_BY_TYPE].items())
|
||||
]
|
||||
if stat_data[REQ_CNT_BY_TYPE]
|
||||
else ["<tr><td colspan='10' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
else ["<tr><td colspan='11' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
)
|
||||
# 按模块分类统计
|
||||
module_rows = "\n".join(
|
||||
|
|
@ -975,11 +987,12 @@ class StatisticOutputTask(AsyncTask):
|
|||
f"<td>{stat_data[STD_TIME_COST_BY_MODULE][module_name]:.1f} 秒</td>"
|
||||
f"<td>{_format_large_number(count / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_MODULE][module_name] / total_replies, html=True) if total_replies > 0 else 'N/A'}</td>"
|
||||
f"<td>{_format_large_number(stat_data[TOTAL_TOK_BY_MODULE][module_name] / count, html=True) if count > 0 else 'N/A'}</td>"
|
||||
f"</tr>"
|
||||
for module_name, count in sorted(stat_data[REQ_CNT_BY_MODULE].items())
|
||||
]
|
||||
if stat_data[REQ_CNT_BY_MODULE]
|
||||
else ["<tr><td colspan='10' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
else ["<tr><td colspan='11' style='text-align: center; color: #999;'>暂无数据</td></tr>"]
|
||||
)
|
||||
|
||||
# 聊天消息统计
|
||||
|
|
@ -1054,7 +1067,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
<h2>按模型分类统计</h2>
|
||||
<div class=\"table-wrap\">
|
||||
<table>
|
||||
<thead><tr><th>模型名称</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th></tr></thead>
|
||||
<thead><tr><th>模型名称</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th><th>每次调用平均Token</th></tr></thead>
|
||||
<tbody>
|
||||
{model_rows}
|
||||
</tbody>
|
||||
|
|
@ -1065,7 +1078,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
<div class=\"table-wrap\">
|
||||
<table>
|
||||
<thead>
|
||||
<tr><th>模块名称</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th></tr>
|
||||
<tr><th>模块名称</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th><th>每次调用平均Token</th></tr>
|
||||
</thead>
|
||||
<tbody>
|
||||
{module_rows}
|
||||
|
|
@ -1077,7 +1090,7 @@ class StatisticOutputTask(AsyncTask):
|
|||
<div class=\"table-wrap\">
|
||||
<table>
|
||||
<thead>
|
||||
<tr><th>请求类型</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th></tr>
|
||||
<tr><th>请求类型</th><th>调用次数</th><th>输入Token</th><th>输出Token</th><th>Token总量</th><th>累计花费</th><th>平均耗时(秒)</th><th>标准差(秒)</th><th>每次回复平均调用次数</th><th>每次回复平均Token数</th><th>每次调用平均Token</th></tr>
|
||||
</thead>
|
||||
<tbody>
|
||||
{type_rows}
|
||||
|
|
|
|||
|
|
@ -260,12 +260,32 @@ class MemoryConfig(ConfigBase):
|
|||
agent_timeout_seconds: float = 120.0
|
||||
"""Agent超时时间（秒）"""
|
||||
|
||||
enable_jargon_detection: bool = True
|
||||
"""记忆检索过程中是否启用黑话识别"""
|
||||
|
||||
global_memory: bool = False
|
||||
"""是否允许记忆检索在聊天记录中进行全局查询（忽略当前chat_id，仅对 search_chat_history 等工具生效）"""
|
||||
|
||||
global_memory_blacklist: list[str] = field(default_factory=lambda: [])
|
||||
"""
|
||||
全局记忆黑名单，当启用全局记忆时，不将特定聊天流纳入检索
|
||||
格式: ["platform:id:type", ...]
|
||||
|
||||
示例:
|
||||
[
|
||||
"qq:1919810:private", # 排除特定私聊
|
||||
"qq:114514:group", # 排除特定群聊
|
||||
]
|
||||
|
||||
说明:
|
||||
- 当启用全局记忆时，黑名单中的聊天流不会被检索
|
||||
- 当在黑名单中的聊天流进行查询时，仅使用该聊天流的本地记忆
|
||||
"""
|
||||
|
||||
planner_question: bool = True
|
||||
"""
|
||||
是否使用 Planner 提供的 question 作为记忆检索问题
|
||||
- True: 当 Planner 在 reply 动作中提供了 question 时，直接使用该问题进行记忆检索，跳过 LLM 生成问题的步骤
|
||||
- False: 沿用旧模式，使用 LLM 生成问题
|
||||
"""
|
||||
|
||||
def __post_init__(self):
|
||||
"""验证配置值"""
|
||||
if self.max_agent_iterations < 1:
|
||||
|
|
|
|||
|
|
@ -912,8 +912,11 @@ class ChatHistorySummarizer:
|
|||
result = _parse_with_quote_fix(extracted_json)
|
||||
|
||||
keywords = result.get("keywords", [])
|
||||
summary = result.get("summary", "无概括")
|
||||
summary = result.get("summary", "")
|
||||
key_point = result.get("key_point", [])
|
||||
|
||||
if not (keywords and summary) and key_point:
|
||||
logger.warning(f"{self.log_prefix} LLM返回的JSON中缺少字段，原文\n{response}")
|
||||
|
||||
# 确保keywords和key_point是列表
|
||||
if isinstance(keywords, str):
|
||||
|
|
|
|||
|
|
@ -2,7 +2,7 @@ import time
|
|||
import json
|
||||
import asyncio
|
||||
import re
|
||||
from typing import List, Dict, Any, Optional, Tuple, Set
|
||||
from typing import List, Dict, Any, Optional, Tuple
|
||||
from src.common.logger import get_logger
|
||||
from src.config.config import global_config, model_config
|
||||
from src.chat.utils.prompt_builder import Prompt, global_prompt_manager
|
||||
|
|
@ -11,7 +11,8 @@ from src.common.database.database_model import ThinkingBack
|
|||
from src.memory_system.retrieval_tools import get_tool_registry, init_all_tools
|
||||
from src.memory_system.memory_utils import parse_questions_json
|
||||
from src.llm_models.payload_content.message import MessageBuilder, RoleType, Message
|
||||
from src.bw_learner.jargon_explainer import match_jargon_from_text, retrieve_concepts_with_jargon
|
||||
from src.chat.message_receive.chat_stream import get_chat_manager
|
||||
from src.bw_learner.jargon_explainer import retrieve_concepts_with_jargon
|
||||
|
||||
logger = get_logger("memory_retrieval")
|
||||
|
||||
|
|
@ -100,6 +101,7 @@ def init_memory_retrieval_prompt():
|
|||
**工具说明：**
|
||||
- 如果涉及过往事件，或者查询某个过去可能提到过的概念，或者某段时间发生的事件。可以使用聊天记录查询工具查询过往事件
|
||||
- 如果涉及人物，可以使用人物信息查询工具查询人物信息
|
||||
- 如果遇到不熟悉的词语、缩写、黑话或网络用语，可以使用query_words工具查询其含义
|
||||
- 如果没有可靠信息，且查询时间充足，或者不确定查询类别，也可以使用lpmm知识库查询，作为辅助信息
|
||||
|
||||
**思考**
|
||||
|
|
@ -202,7 +204,6 @@ async def _react_agent_solve_question(
|
|||
max_iterations: int = 5,
|
||||
timeout: float = 30.0,
|
||||
initial_info: str = "",
|
||||
initial_jargon_concepts: Optional[List[str]] = None,
|
||||
) -> Tuple[bool, str, List[Dict[str, Any]], bool]:
|
||||
"""使用ReAct架构的Agent来解决问题
|
||||
|
||||
|
|
@ -211,28 +212,29 @@ async def _react_agent_solve_question(
|
|||
chat_id: 聊天ID
|
||||
max_iterations: 最大迭代次数
|
||||
timeout: 超时时间（秒）
|
||||
initial_info: 初始信息（如概念检索结果），将作为collected_info的初始值
|
||||
initial_jargon_concepts: 预先已解析过的黑话列表，避免重复解释
|
||||
initial_info: 初始信息，将作为collected_info的初始值
|
||||
|
||||
Returns:
|
||||
Tuple[bool, str, List[Dict[str, Any]], bool]: (是否找到答案, 答案内容, 思考步骤列表, 是否超时)
|
||||
"""
|
||||
start_time = time.time()
|
||||
collected_info = initial_info if initial_info else ""
|
||||
enable_jargon_detection = global_config.memory.enable_jargon_detection
|
||||
seen_jargon_concepts: Set[str] = set()
|
||||
if enable_jargon_detection and initial_jargon_concepts:
|
||||
for concept in initial_jargon_concepts:
|
||||
concept = (concept or "").strip()
|
||||
if concept:
|
||||
seen_jargon_concepts.add(concept)
|
||||
# 构造日志前缀：[聊天流名称]，用于在日志中标识聊天流
|
||||
try:
|
||||
chat_name = get_chat_manager().get_stream_name(chat_id) or chat_id
|
||||
except Exception:
|
||||
chat_name = chat_id
|
||||
react_log_prefix = f"[{chat_name}] "
|
||||
thinking_steps = []
|
||||
is_timeout = False
|
||||
conversation_messages: List[Message] = []
|
||||
first_head_prompt: Optional[str] = None # 保存第一次使用的head_prompt（用于日志显示）
|
||||
last_tool_name: Optional[str] = None # 记录最后一次使用的工具名称
|
||||
|
||||
# 正常迭代：max_iterations 次（最终评估单独处理，不算在迭代中）
|
||||
for iteration in range(max_iterations):
|
||||
# 使用 while 循环，支持额外迭代
|
||||
iteration = 0
|
||||
max_iterations_with_extra = max_iterations
|
||||
while iteration < max_iterations_with_extra:
|
||||
# 检查超时
|
||||
if time.time() - start_time > timeout:
|
||||
logger.warning(f"ReAct Agent超时，已迭代{iteration}次")
|
||||
|
|
@ -475,7 +477,7 @@ async def _react_agent_solve_question(
|
|||
step["observations"] = ["检测到finish_search文本格式调用，找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(
|
||||
f"ReAct Agent 第 {iteration + 1} 次迭代 通过finish_search文本格式找到关于问题{question}的答案: {parsed_answer}"
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 通过finish_search文本格式找到关于问题{question}的答案: {parsed_answer}"
|
||||
)
|
||||
|
||||
_log_conversation_messages(
|
||||
|
|
@ -488,7 +490,7 @@ async def _react_agent_solve_question(
|
|||
else:
|
||||
# found_answer为True但没有提供answer，视为错误，继续迭代
|
||||
logger.warning(
|
||||
f"ReAct Agent 第 {iteration + 1} 次迭代 finish_search文本格式found_answer为True但未提供answer"
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 finish_search文本格式found_answer为True但未提供answer"
|
||||
)
|
||||
else:
|
||||
# 未找到答案，直接退出查询
|
||||
|
|
@ -497,7 +499,9 @@ async def _react_agent_solve_question(
|
|||
)
|
||||
step["observations"] = ["检测到finish_search文本格式调用，未找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 通过finish_search文本格式判断未找到答案")
|
||||
logger.info(
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 通过finish_search文本格式判断未找到答案"
|
||||
)
|
||||
|
||||
_log_conversation_messages(
|
||||
conversation_messages,
|
||||
|
|
@ -509,10 +513,12 @@ async def _react_agent_solve_question(
|
|||
|
||||
# 如果没有检测到finish_search格式，记录思考过程，继续下一轮迭代
|
||||
step["observations"] = [f"思考完成，但未调用工具。响应: {response}"]
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 思考完成但未调用工具: {response}")
|
||||
logger.info(
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 思考完成但未调用工具: {response}"
|
||||
)
|
||||
collected_info += f"思考: {response}"
|
||||
else:
|
||||
logger.warning(f"ReAct Agent 第 {iteration + 1} 次迭代 无工具调用且无响应")
|
||||
logger.warning(f"{react_log_prefix}第 {iteration + 1} 次迭代 无工具调用且无响应")
|
||||
step["observations"] = ["无响应且无工具调用"]
|
||||
thinking_steps.append(step)
|
||||
continue
|
||||
|
|
@ -541,7 +547,7 @@ async def _react_agent_solve_question(
|
|||
step["observations"] = ["检测到finish_search工具调用，找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(
|
||||
f"ReAct Agent 第 {iteration + 1} 次迭代 通过finish_search工具找到关于问题{question}的答案: {finish_search_answer}"
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 通过finish_search工具找到关于问题{question}的答案: {finish_search_answer}"
|
||||
)
|
||||
|
||||
_log_conversation_messages(
|
||||
|
|
@ -554,14 +560,16 @@ async def _react_agent_solve_question(
|
|||
else:
|
||||
# found_answer为True但没有提供answer，视为错误
|
||||
logger.warning(
|
||||
f"ReAct Agent 第 {iteration + 1} 次迭代 finish_search工具found_answer为True但未提供answer"
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 finish_search工具found_answer为True但未提供answer"
|
||||
)
|
||||
else:
|
||||
# 未找到答案，直接退出查询
|
||||
step["actions"].append({"action_type": "finish_search", "action_params": {"found_answer": False}})
|
||||
step["observations"] = ["检测到finish_search工具调用，未找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 通过finish_search工具判断未找到答案")
|
||||
logger.info(
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 通过finish_search工具判断未找到答案"
|
||||
)
|
||||
|
||||
_log_conversation_messages(
|
||||
conversation_messages,
|
||||
|
|
@ -578,13 +586,16 @@ async def _react_agent_solve_question(
|
|||
tool_args = tool_call.args or {}
|
||||
|
||||
logger.debug(
|
||||
f"ReAct Agent 第 {iteration + 1} 次迭代 工具调用 {i + 1}/{len(tool_calls)}: {tool_name}({tool_args})"
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 工具调用 {i + 1}/{len(tool_calls)}: {tool_name}({tool_args})"
|
||||
)
|
||||
|
||||
# 跳过finish_search工具调用（已经在上面处理过了）
|
||||
if tool_name == "finish_search":
|
||||
continue
|
||||
|
||||
# 记录最后一次使用的工具名称（用于判断是否需要额外迭代）
|
||||
last_tool_name = tool_name
|
||||
|
||||
# 普通工具调用
|
||||
tool = tool_registry.get_tool(tool_name)
|
||||
if tool:
|
||||
|
|
@ -604,14 +615,18 @@ async def _react_agent_solve_question(
|
|||
return f"查询{tool_name_str}({param_str})的结果：{observation}"
|
||||
except Exception as e:
|
||||
error_msg = f"工具执行失败: {str(e)}"
|
||||
logger.error(f"ReAct Agent 第 {iter_num + 1} 次迭代 工具 {tool_name_str} {error_msg}")
|
||||
logger.error(
|
||||
f"{react_log_prefix}第 {iter_num + 1} 次迭代 工具 {tool_name_str} {error_msg}"
|
||||
)
|
||||
return f"查询{tool_name_str}失败: {error_msg}"
|
||||
|
||||
tool_tasks.append(execute_single_tool(tool, tool_params, tool_name, iteration))
|
||||
step["actions"].append({"action_type": tool_name, "action_params": tool_args})
|
||||
else:
|
||||
error_msg = f"未知的工具类型: {tool_name}"
|
||||
logger.warning(f"ReAct Agent 第 {iteration + 1} 次迭代 工具 {i + 1}/{len(tool_calls)} {error_msg}")
|
||||
logger.warning(
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 工具 {i + 1}/{len(tool_calls)} {error_msg}"
|
||||
)
|
||||
tool_tasks.append(asyncio.create_task(asyncio.sleep(0, result=f"查询{tool_name}失败: {error_msg}")))
|
||||
|
||||
# 并行执行所有工具
|
||||
|
|
@ -622,31 +637,16 @@ async def _react_agent_solve_question(
|
|||
for i, (tool_call_item, observation) in enumerate(zip(tool_calls, observations, strict=False)):
|
||||
if isinstance(observation, Exception):
|
||||
observation = f"工具执行异常: {str(observation)}"
|
||||
logger.error(f"ReAct Agent 第 {iteration + 1} 次迭代 工具 {i + 1} 执行异常: {observation}")
|
||||
logger.error(
|
||||
f"{react_log_prefix}第 {iteration + 1} 次迭代 工具 {i + 1} 执行异常: {observation}"
|
||||
)
|
||||
|
||||
observation_text = observation if isinstance(observation, str) else str(observation)
|
||||
stripped_observation = observation_text.strip()
|
||||
step["observations"].append(observation_text)
|
||||
collected_info += f"\n{observation_text}\n"
|
||||
if stripped_observation:
|
||||
# 检查工具输出中是否有新的jargon，如果有则追加到工具结果中
|
||||
if enable_jargon_detection:
|
||||
jargon_concepts = match_jargon_from_text(stripped_observation, chat_id)
|
||||
if jargon_concepts:
|
||||
new_concepts = []
|
||||
for concept in jargon_concepts:
|
||||
normalized_concept = concept.strip()
|
||||
if normalized_concept and normalized_concept not in seen_jargon_concepts:
|
||||
new_concepts.append(normalized_concept)
|
||||
seen_jargon_concepts.add(normalized_concept)
|
||||
if new_concepts:
|
||||
jargon_info = await retrieve_concepts_with_jargon(new_concepts, chat_id)
|
||||
if jargon_info:
|
||||
# 将jargon查询结果追加到工具结果中
|
||||
observation_text += f"\n\n{jargon_info}"
|
||||
collected_info += f"\n{jargon_info}\n"
|
||||
logger.info(f"工具输出触发黑话解析: {new_concepts}")
|
||||
|
||||
# 不再自动检测工具输出中的jargon，改为通过 query_words 工具主动查询
|
||||
tool_builder = MessageBuilder()
|
||||
tool_builder.set_role(RoleType.Tool)
|
||||
tool_builder.add_text_content(observation_text)
|
||||
|
|
@ -655,15 +655,24 @@ async def _react_agent_solve_question(
|
|||
|
||||
thinking_steps.append(step)
|
||||
|
||||
# 检查是否需要额外迭代：如果最后一次使用的工具是 search_chat_history 且达到最大迭代次数，额外增加一回合
|
||||
if iteration + 1 >= max_iterations and last_tool_name == "search_chat_history" and not is_timeout:
|
||||
max_iterations_with_extra = max_iterations + 1
|
||||
logger.info(
|
||||
f"{react_log_prefix}达到最大迭代次数（已迭代{iteration + 1}次），最后一次使用工具为 search_chat_history，额外增加一回合尝试"
|
||||
)
|
||||
|
||||
iteration += 1
|
||||
|
||||
# 正常迭代结束后，如果达到最大迭代次数或超时，执行最终评估
|
||||
# 最终评估单独处理，不算在迭代中
|
||||
should_do_final_evaluation = False
|
||||
if is_timeout:
|
||||
should_do_final_evaluation = True
|
||||
logger.warning(f"ReAct Agent超时，已迭代{iteration + 1}次，进入最终评估")
|
||||
elif iteration + 1 >= max_iterations:
|
||||
logger.warning(f"{react_log_prefix}超时，已迭代{iteration}次，进入最终评估")
|
||||
elif iteration >= max_iterations:
|
||||
should_do_final_evaluation = True
|
||||
logger.info(f"ReAct Agent达到最大迭代次数（已迭代{iteration + 1}次），进入最终评估")
|
||||
logger.info(f"{react_log_prefix}达到最大迭代次数（已迭代{iteration}次），进入最终评估")
|
||||
|
||||
if should_do_final_evaluation:
|
||||
# 获取必要变量用于最终评估
|
||||
|
|
@ -766,8 +775,8 @@ async def _react_agent_solve_question(
|
|||
return False, "最终评估阶段LLM调用失败", thinking_steps, is_timeout
|
||||
|
||||
if global_config.debug.show_memory_prompt:
|
||||
logger.info(f"ReAct Agent 最终评估Prompt: {evaluation_prompt}")
|
||||
logger.info(f"ReAct Agent 最终评估响应: {eval_response}")
|
||||
logger.info(f"{react_log_prefix}最终评估Prompt: {evaluation_prompt}")
|
||||
logger.info(f"{react_log_prefix}最终评估响应: {eval_response}")
|
||||
|
||||
# 从最终评估响应中提取found_answer或not_enough_info
|
||||
found_answer_content = None
|
||||
|
|
@ -998,7 +1007,6 @@ async def _process_single_question(
|
|||
chat_id: str,
|
||||
context: str,
|
||||
initial_info: str = "",
|
||||
initial_jargon_concepts: Optional[List[str]] = None,
|
||||
max_iterations: Optional[int] = None,
|
||||
) -> Optional[str]:
|
||||
"""处理单个问题的查询
|
||||
|
|
@ -1007,8 +1015,8 @@ async def _process_single_question(
|
|||
question: 要查询的问题
|
||||
chat_id: 聊天ID
|
||||
context: 上下文信息
|
||||
initial_info: 初始信息（如概念检索结果），将传递给ReAct Agent
|
||||
initial_jargon_concepts: 已经处理过的黑话概念列表，用于ReAct阶段的去重
|
||||
initial_info: 初始信息，将传递给ReAct Agent
|
||||
max_iterations: 最大迭代次数
|
||||
|
||||
Returns:
|
||||
Optional[str]: 如果找到答案，返回格式化的结果字符串，否则返回None
|
||||
|
|
@ -1022,8 +1030,6 @@ async def _process_single_question(
|
|||
# 直接使用ReAct Agent查询（不再从thinking_back获取缓存）
|
||||
# logger.info(f"使用ReAct Agent查询，问题: {question[:50]}...")
|
||||
|
||||
jargon_concepts_for_agent = initial_jargon_concepts if global_config.memory.enable_jargon_detection else None
|
||||
|
||||
# 如果未指定max_iterations，使用配置的默认值
|
||||
if max_iterations is None:
|
||||
max_iterations = global_config.memory.max_agent_iterations
|
||||
|
|
@ -1034,7 +1040,6 @@ async def _process_single_question(
|
|||
max_iterations=max_iterations,
|
||||
timeout=global_config.memory.agent_timeout_seconds,
|
||||
initial_info=question_initial_info,
|
||||
initial_jargon_concepts=jargon_concepts_for_agent,
|
||||
)
|
||||
|
||||
# 存储查询历史到数据库（超时时不存储）
|
||||
|
|
@ -1062,6 +1067,8 @@ async def build_memory_retrieval_prompt(
|
|||
target: str,
|
||||
chat_stream,
|
||||
think_level: int = 1,
|
||||
unknown_words: Optional[List[str]] = None,
|
||||
question: Optional[str] = None,
|
||||
) -> str:
|
||||
"""构建记忆检索提示
|
||||
使用两段式查询：第一步生成问题，第二步使用ReAct Agent查询答案
|
||||
|
|
@ -1071,14 +1078,33 @@ async def build_memory_retrieval_prompt(
|
|||
sender: 发送者名称
|
||||
target: 目标消息内容
|
||||
chat_stream: 聊天流对象
|
||||
tool_executor: 工具执行器（保留参数以兼容接口）
|
||||
think_level: 思考深度等级
|
||||
unknown_words: Planner 提供的未知词语列表，优先使用此列表而不是从聊天记录匹配
|
||||
question: Planner 提供的问题，当 planner_question 配置开启时，直接使用此问题进行检索
|
||||
|
||||
Returns:
|
||||
str: 记忆检索结果字符串
|
||||
"""
|
||||
start_time = time.time()
|
||||
|
||||
logger.info(f"检测是否需要回忆，元消息：{message[:30]}...，消息长度: {len(message)}")
|
||||
# 构造日志前缀：[聊天流名称]，用于在日志中标识聊天流（优先群名称/用户昵称）
|
||||
try:
|
||||
group_info = chat_stream.group_info
|
||||
user_info = chat_stream.user_info
|
||||
# 群聊优先使用群名称
|
||||
if group_info is not None and getattr(group_info, "group_name", None):
|
||||
stream_name = group_info.group_name.strip() or str(group_info.group_id)
|
||||
# 私聊使用用户昵称
|
||||
elif user_info is not None and getattr(user_info, "user_nickname", None):
|
||||
stream_name = user_info.user_nickname.strip() or str(user_info.user_id)
|
||||
# 兜底使用 stream_id
|
||||
else:
|
||||
stream_name = chat_stream.stream_id
|
||||
except Exception:
|
||||
stream_name = chat_stream.stream_id
|
||||
log_prefix = f"[{stream_name}] " if stream_name else ""
|
||||
|
||||
logger.info(f"{log_prefix}检测是否需要回忆，元消息：{message[:30]}...，消息长度: {len(message)}")
|
||||
try:
|
||||
time_now = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
|
||||
bot_name = global_config.bot.nickname
|
||||
|
|
@ -1089,63 +1115,78 @@ async def build_memory_retrieval_prompt(
|
|||
if not recent_query_history:
|
||||
recent_query_history = "最近没有查询记录。"
|
||||
|
||||
# 第一步：生成问题
|
||||
question_prompt = await global_prompt_manager.format_prompt(
|
||||
"memory_retrieval_question_prompt",
|
||||
bot_name=bot_name,
|
||||
time_now=time_now,
|
||||
chat_history=message,
|
||||
recent_query_history=recent_query_history,
|
||||
sender=sender,
|
||||
target_message=target,
|
||||
)
|
||||
|
||||
success, response, reasoning_content, model_name = await llm_api.generate_with_model(
|
||||
question_prompt,
|
||||
model_config=model_config.model_task_config.tool_use,
|
||||
request_type="memory.question",
|
||||
)
|
||||
|
||||
if global_config.debug.show_memory_prompt:
|
||||
logger.info(f"记忆检索问题生成提示词: {question_prompt}")
|
||||
# logger.info(f"记忆检索问题生成响应: {response}")
|
||||
|
||||
if not success:
|
||||
logger.error(f"LLM生成问题失败: {response}")
|
||||
return ""
|
||||
|
||||
# 解析概念列表和问题列表
|
||||
_, questions = parse_questions_json(response)
|
||||
if questions:
|
||||
logger.info(f"解析到 {len(questions)} 个问题: {questions}")
|
||||
|
||||
enable_jargon_detection = global_config.memory.enable_jargon_detection
|
||||
concepts: List[str] = []
|
||||
|
||||
if enable_jargon_detection:
|
||||
# 使用匹配逻辑自动识别聊天中的黑话概念
|
||||
concepts = match_jargon_from_text(message, chat_id)
|
||||
if concepts:
|
||||
logger.info(f"黑话匹配命中 {len(concepts)} 个概念: {concepts}")
|
||||
# 第一步：生成问题或使用 Planner 提供的问题
|
||||
questions = []
|
||||
|
||||
# 如果 planner_question 配置开启，只使用 Planner 提供的问题，不使用旧模式
|
||||
if global_config.memory.planner_question:
|
||||
if question and isinstance(question, str) and question.strip():
|
||||
# 清理和验证 question
|
||||
cleaned_question = question.strip()
|
||||
questions = [cleaned_question]
|
||||
logger.info(f"{log_prefix}使用 Planner 提供的 question: {cleaned_question}")
|
||||
else:
|
||||
logger.debug("黑话匹配未命中任何概念")
|
||||
# planner_question 开启但没有提供 question，跳过记忆检索
|
||||
logger.debug(f"{log_prefix}planner_question 已开启但未提供 question，跳过记忆检索")
|
||||
end_time = time.time()
|
||||
logger.info(f"{log_prefix}无当次查询，不返回任何结果，耗时: {(end_time - start_time):.3f}秒")
|
||||
return ""
|
||||
else:
|
||||
logger.debug("已禁用记忆检索中的黑话识别")
|
||||
# planner_question 关闭，使用旧模式：LLM 生成问题
|
||||
question_prompt = await global_prompt_manager.format_prompt(
|
||||
"memory_retrieval_question_prompt",
|
||||
bot_name=bot_name,
|
||||
time_now=time_now,
|
||||
chat_history=message,
|
||||
recent_query_history=recent_query_history,
|
||||
sender=sender,
|
||||
target_message=target,
|
||||
)
|
||||
|
||||
# 对匹配到的概念进行jargon检索，作为初始信息
|
||||
success, response, reasoning_content, model_name = await llm_api.generate_with_model(
|
||||
question_prompt,
|
||||
model_config=model_config.model_task_config.tool_use,
|
||||
request_type="memory.question",
|
||||
)
|
||||
|
||||
if global_config.debug.show_memory_prompt:
|
||||
logger.info(f"{log_prefix}记忆检索问题生成提示词: {question_prompt}")
|
||||
# logger.info(f"记忆检索问题生成响应: {response}")
|
||||
|
||||
if not success:
|
||||
logger.error(f"{log_prefix}LLM生成问题失败: {response}")
|
||||
return ""
|
||||
|
||||
# 解析概念列表和问题列表
|
||||
_, questions = parse_questions_json(response)
|
||||
if questions:
|
||||
logger.info(f"{log_prefix}解析到 {len(questions)} 个问题: {questions}")
|
||||
|
||||
# 初始阶段：使用 Planner 提供的 unknown_words 进行检索（如果提供）
|
||||
initial_info = ""
|
||||
if enable_jargon_detection and concepts:
|
||||
concept_info = await retrieve_concepts_with_jargon(concepts, chat_id)
|
||||
if concept_info:
|
||||
initial_info += concept_info
|
||||
logger.debug(f"概念检索完成，结果: {concept_info}")
|
||||
else:
|
||||
logger.debug("概念检索未找到任何结果")
|
||||
if unknown_words and len(unknown_words) > 0:
|
||||
# 清理和去重 unknown_words
|
||||
cleaned_concepts = []
|
||||
for word in unknown_words:
|
||||
if isinstance(word, str):
|
||||
cleaned = word.strip()
|
||||
if cleaned:
|
||||
cleaned_concepts.append(cleaned)
|
||||
if cleaned_concepts:
|
||||
# 对匹配到的概念进行jargon检索，作为初始信息
|
||||
concept_info = await retrieve_concepts_with_jargon(cleaned_concepts, chat_id)
|
||||
if concept_info:
|
||||
initial_info += concept_info
|
||||
logger.info(
|
||||
f"{log_prefix}使用 Planner 提供的 unknown_words，共 {len(cleaned_concepts)} 个概念，检索结果: {concept_info[:100]}..."
|
||||
)
|
||||
else:
|
||||
logger.debug(f"{log_prefix}unknown_words 检索未找到任何结果")
|
||||
|
||||
if not questions:
|
||||
logger.debug("模型认为不需要检索记忆或解析失败，不返回任何查询结果")
|
||||
logger.debug(f"{log_prefix}模型认为不需要检索记忆或解析失败，不返回任何查询结果")
|
||||
end_time = time.time()
|
||||
logger.info(f"无当次查询，不返回任何结果，耗时: {(end_time - start_time):.3f}秒")
|
||||
logger.info(f"{log_prefix}无当次查询，不返回任何结果，耗时: {(end_time - start_time):.3f}秒")
|
||||
return ""
|
||||
|
||||
# 第二步：并行处理所有问题（使用配置的最大迭代次数和超时时间）
|
||||
|
|
@ -1157,17 +1198,16 @@ async def build_memory_retrieval_prompt(
|
|||
max_iterations = base_max_iterations
|
||||
timeout_seconds = global_config.memory.agent_timeout_seconds
|
||||
logger.debug(
|
||||
f"问题数量: {len(questions)}，think_level={think_level}，设置最大迭代次数: {max_iterations}（基础值: {base_max_iterations}），超时时间: {timeout_seconds}秒"
|
||||
f"{log_prefix}问题数量: {len(questions)}，think_level={think_level}，设置最大迭代次数: {max_iterations}（基础值: {base_max_iterations}），超时时间: {timeout_seconds}秒"
|
||||
)
|
||||
|
||||
# 并行处理所有问题，将概念检索结果作为初始信息传递
|
||||
# 并行处理所有问题
|
||||
question_tasks = [
|
||||
_process_single_question(
|
||||
question=question,
|
||||
chat_id=chat_id,
|
||||
context=message,
|
||||
initial_info=initial_info,
|
||||
initial_jargon_concepts=concepts if enable_jargon_detection else None,
|
||||
max_iterations=max_iterations,
|
||||
)
|
||||
for question in questions
|
||||
|
|
@ -1180,7 +1220,7 @@ async def build_memory_retrieval_prompt(
|
|||
question_results: List[str] = []
|
||||
for i, result in enumerate(results):
|
||||
if isinstance(result, Exception):
|
||||
logger.error(f"处理问题 '{questions[i]}' 时发生异常: {result}")
|
||||
logger.error(f"{log_prefix}处理问题 '{questions[i]}' 时发生异常: {result}")
|
||||
elif result is not None:
|
||||
question_results.append(result)
|
||||
|
||||
|
|
@ -1216,14 +1256,14 @@ async def build_memory_retrieval_prompt(
|
|||
current_count = len(question_results)
|
||||
cached_count = len(all_results) - current_count
|
||||
logger.info(
|
||||
f"记忆检索成功，耗时: {(end_time - start_time):.3f}秒，"
|
||||
f"{log_prefix}记忆检索成功，耗时: {(end_time - start_time):.3f}秒，"
|
||||
f"当前查询 {current_count} 条记忆，缓存 {cached_count} 条记忆，共 {len(all_results)} 条记忆"
|
||||
)
|
||||
return f"你回忆起了以下信息：\n{retrieved_memory}\n如果与回复内容相关，可以参考这些回忆的信息。\n"
|
||||
else:
|
||||
logger.debug("所有问题均未找到答案，且无缓存答案")
|
||||
logger.debug(f"{log_prefix}所有问题均未找到答案，且无缓存答案")
|
||||
return ""
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"记忆检索时发生异常: {str(e)}")
|
||||
logger.error(f"{log_prefix}记忆检索时发生异常: {str(e)}")
|
||||
return ""
|
||||
|
|
|
|||
|
|
@ -14,6 +14,7 @@ from .tool_registry import (
|
|||
from .query_chat_history import register_tool as register_query_chat_history
|
||||
from .query_lpmm_knowledge import register_tool as register_lpmm_knowledge
|
||||
from .query_person_info import register_tool as register_query_person_info
|
||||
from .query_words import register_tool as register_query_words
|
||||
from .found_answer import register_tool as register_finish_search
|
||||
from src.config.config import global_config
|
||||
|
||||
|
|
@ -22,6 +23,7 @@ def init_all_tools():
|
|||
"""初始化并注册所有记忆检索工具"""
|
||||
register_query_chat_history()
|
||||
register_query_person_info()
|
||||
register_query_words() # 注册query_words工具
|
||||
register_finish_search() # 注册finish_search工具
|
||||
|
||||
if global_config.lpmm_knowledge.lpmm_mode == "agent":
|
||||
|
|
|
|||
|
|
@ -4,7 +4,7 @@
|
|||
"""
|
||||
|
||||
import json
|
||||
from typing import Optional
|
||||
from typing import Optional, Set
|
||||
from datetime import datetime
|
||||
|
||||
from src.common.logger import get_logger
|
||||
|
|
@ -16,6 +16,72 @@ from .tool_registry import register_memory_retrieval_tool
|
|||
logger = get_logger("memory_retrieval_tools")
|
||||
|
||||
|
||||
def _parse_blacklist_to_chat_ids(blacklist: list[str]) -> Set[str]:
|
||||
"""将黑名单配置（platform:id:type格式）转换为chat_id集合
|
||||
|
||||
Args:
|
||||
blacklist: 黑名单配置列表，格式为 ["platform:id:type", ...]
|
||||
|
||||
Returns:
|
||||
Set[str]: chat_id集合
|
||||
"""
|
||||
chat_ids = set()
|
||||
if not blacklist:
|
||||
return chat_ids
|
||||
|
||||
try:
|
||||
from src.chat.message_receive.chat_stream import get_chat_manager
|
||||
|
||||
chat_manager = get_chat_manager()
|
||||
for blacklist_item in blacklist:
|
||||
if not isinstance(blacklist_item, str):
|
||||
continue
|
||||
|
||||
try:
|
||||
parts = blacklist_item.split(":")
|
||||
if len(parts) != 3:
|
||||
logger.warning(f"黑名单配置格式错误，应为 platform:id:type，实际: {blacklist_item}")
|
||||
continue
|
||||
|
||||
platform = parts[0]
|
||||
id_str = parts[1]
|
||||
stream_type = parts[2]
|
||||
|
||||
# 判断是否为群聊
|
||||
is_group = stream_type == "group"
|
||||
|
||||
# 转换为chat_id
|
||||
chat_id = chat_manager.get_stream_id(platform, str(id_str), is_group=is_group)
|
||||
if chat_id:
|
||||
chat_ids.add(chat_id)
|
||||
else:
|
||||
logger.warning(f"无法将黑名单配置转换为chat_id: {blacklist_item}")
|
||||
except Exception as e:
|
||||
logger.warning(f"解析黑名单配置失败: {blacklist_item}, 错误: {e}")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"初始化黑名单chat_id集合失败: {e}")
|
||||
|
||||
return chat_ids
|
||||
|
||||
|
||||
def _is_chat_id_in_blacklist(chat_id: str) -> bool:
|
||||
"""检查chat_id是否在全局记忆黑名单中
|
||||
|
||||
Args:
|
||||
chat_id: 要检查的chat_id
|
||||
|
||||
Returns:
|
||||
bool: 如果chat_id在黑名单中返回True，否则返回False
|
||||
"""
|
||||
blacklist = getattr(global_config.memory, "global_memory_blacklist", [])
|
||||
if not blacklist:
|
||||
return False
|
||||
|
||||
blacklist_chat_ids = _parse_blacklist_to_chat_ids(blacklist)
|
||||
return chat_id in blacklist_chat_ids
|
||||
|
||||
|
||||
async def search_chat_history(chat_id: str, keyword: Optional[str] = None, participant: Optional[str] = None) -> str:
|
||||
"""根据关键词或参与人查询记忆，返回匹配的记忆id、记忆标题theme和关键词keywords
|
||||
|
||||
|
|
@ -33,17 +99,34 @@ async def search_chat_history(chat_id: str, keyword: Optional[str] = None, parti
|
|||
return "未指定查询参数（需要提供keyword或participant之一）"
|
||||
|
||||
# 构建查询条件
|
||||
# 检查当前chat_id是否在黑名单中
|
||||
is_current_chat_in_blacklist = _is_chat_id_in_blacklist(chat_id)
|
||||
|
||||
# 根据配置决定是否限制在当前 chat_id 内查询
|
||||
use_global_search = global_config.memory.global_memory
|
||||
# 如果当前chat_id在黑名单中，强制使用本地查询
|
||||
use_global_search = global_config.memory.global_memory and not is_current_chat_in_blacklist
|
||||
|
||||
if use_global_search:
|
||||
# 全局查询所有聊天记录
|
||||
query = ChatHistory.select()
|
||||
logger.debug(
|
||||
f"search_chat_history 启用全局查询模式，忽略 chat_id 过滤，keyword={keyword}, participant={participant}"
|
||||
)
|
||||
# 全局查询所有聊天记录，但排除黑名单中的聊天流
|
||||
blacklist_chat_ids = _parse_blacklist_to_chat_ids(global_config.memory.global_memory_blacklist)
|
||||
if blacklist_chat_ids:
|
||||
# 排除黑名单中的chat_id
|
||||
query = ChatHistory.select().where(~(ChatHistory.chat_id.in_(blacklist_chat_ids)))
|
||||
logger.debug(
|
||||
f"search_chat_history 启用全局查询模式（排除黑名单 {len(blacklist_chat_ids)} 个聊天流），keyword={keyword}, participant={participant}"
|
||||
)
|
||||
else:
|
||||
# 没有黑名单，查询所有
|
||||
query = ChatHistory.select()
|
||||
logger.debug(
|
||||
f"search_chat_history 启用全局查询模式，忽略 chat_id 过滤，keyword={keyword}, participant={participant}"
|
||||
)
|
||||
else:
|
||||
# 仅在当前聊天流内查询
|
||||
if is_current_chat_in_blacklist:
|
||||
logger.debug(
|
||||
f"search_chat_history 当前聊天流在黑名单中，强制使用本地查询，chat_id={chat_id}, keyword={keyword}, participant={participant}"
|
||||
)
|
||||
query = ChatHistory.select().where(ChatHistory.chat_id == chat_id)
|
||||
|
||||
# 执行查询
|
||||
|
|
|
|||
|
|
@ -0,0 +1,80 @@
|
|||
"""
|
||||
查询黑话/概念含义 - 工具实现
|
||||
用于在记忆检索过程中主动查询未知词语或黑话的含义
|
||||
"""
|
||||
|
||||
from typing import List, Optional
|
||||
from src.common.logger import get_logger
|
||||
from src.bw_learner.jargon_explainer import retrieve_concepts_with_jargon
|
||||
from .tool_registry import register_memory_retrieval_tool
|
||||
|
||||
logger = get_logger("memory_retrieval_tools")
|
||||
|
||||
|
||||
async def query_words(chat_id: str, words: str) -> str:
|
||||
"""查询词语或黑话的含义
|
||||
|
||||
Args:
|
||||
chat_id: 聊天ID
|
||||
words: 要查询的词语，可以是单个词语或多个词语（用逗号、空格等分隔）
|
||||
|
||||
Returns:
|
||||
str: 查询结果，包含词语的含义解释
|
||||
"""
|
||||
try:
|
||||
if not words or not words.strip():
|
||||
return "未提供要查询的词语"
|
||||
|
||||
# 解析词语列表（支持逗号、空格等分隔符）
|
||||
words_list = []
|
||||
for separator in [",", "，", " ", "\n", "\t"]:
|
||||
if separator in words:
|
||||
words_list = [w.strip() for w in words.split(separator) if w.strip()]
|
||||
break
|
||||
|
||||
# 如果没有找到分隔符，整个字符串作为一个词语
|
||||
if not words_list:
|
||||
words_list = [words.strip()]
|
||||
|
||||
# 去重
|
||||
unique_words = []
|
||||
seen = set()
|
||||
for word in words_list:
|
||||
if word and word not in seen:
|
||||
unique_words.append(word)
|
||||
seen.add(word)
|
||||
|
||||
if not unique_words:
|
||||
return "未提供有效的词语"
|
||||
|
||||
logger.info(f"查询词语含义: {unique_words}")
|
||||
|
||||
# 调用检索函数
|
||||
result = await retrieve_concepts_with_jargon(unique_words, chat_id)
|
||||
|
||||
if result:
|
||||
return result
|
||||
else:
|
||||
return f"未找到词语 '{', '.join(unique_words)}' 的含义或黑话解释"
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"查询词语含义失败: {e}")
|
||||
return f"查询失败: {str(e)}"
|
||||
|
||||
|
||||
def register_tool():
|
||||
"""注册工具"""
|
||||
register_memory_retrieval_tool(
|
||||
name="query_words",
|
||||
description="查询词语或黑话的含义。当遇到不熟悉的词语、缩写、黑话或网络用语时，可以使用此工具查询其含义。支持查询单个或多个词语（用逗号、空格等分隔）。",
|
||||
parameters=[
|
||||
{
|
||||
"name": "words",
|
||||
"type": "string",
|
||||
"description": "要查询的词语，可以是单个词语或多个词语（用逗号、空格等分隔，如：'YYDS' 或 'YYDS,内卷,996'）",
|
||||
"required": True,
|
||||
},
|
||||
],
|
||||
execute_func=query_words,
|
||||
)
|
||||
|
||||
|
|
@ -1,5 +1,5 @@
|
|||
[inner]
|
||||
version = "7.2.5"
|
||||
version = "7.2.8"
|
||||
|
||||
#----以下是给开发人员阅读的，如果你只是部署了麦麦，不需要阅读----
|
||||
# 如果你想要修改配置文件，请递增version的值
|
||||
|
|
@ -122,10 +122,13 @@ talk_value_rules = [
|
|||
]
|
||||
|
||||
[memory]
|
||||
max_agent_iterations = 3 # 记忆思考深度（最低为1）
|
||||
agent_timeout_seconds = 200.0 # 最长回忆时间（秒）
|
||||
enable_jargon_detection = true # 记忆检索过程中是否启用黑话识别
|
||||
max_agent_iterations = 5 # 记忆思考深度（最低为1）
|
||||
agent_timeout_seconds = 180.0 # 最长回忆时间（秒）
|
||||
global_memory = false # 是否允许记忆检索进行全局查询
|
||||
global_memory_blacklist = [
|
||||
|
||||
] # 全局记忆黑名单，当启用全局记忆时，不将特定聊天流纳入检索。格式: ["platform:id:type", ...]，例如: ["qq:1919810:private", "qq:114514:group"]
|
||||
planner_question = true # 是否使用 Planner 提供的 question 作为记忆检索问题。开启后，当 Planner 在 reply 动作中提供了 question 时，直接使用该问题进行记忆检索，跳过 LLM 生成问题的步骤；关闭后沿用旧模式，使用 LLM 生成问题
|
||||
|
||||
[dream]
|
||||
interval_minutes = 60 # 做梦时间间隔（分钟），默认30分钟
|
||||
|
|
|
|||
Loading…
Reference in New Issue