mirror of https://github.com/Mai-with-u/MaiBot.git
better：优化记忆提取能力
SengokuCola
parent
82a87f4926
commit
2d6eba7da1
|
|
@ -112,10 +112,10 @@ class FrequencyControl:
|
|||
if len(response) < 20:
|
||||
if "过于频繁" in response:
|
||||
logger.info(f"频率调整: 过于频繁，调整值到{final_value_by_api}")
|
||||
self.talk_frequency_adjust = max(0.1, min(3.0, self.talk_frequency_adjust * 0.8))
|
||||
self.talk_frequency_adjust = max(0.1, min(1.5, self.talk_frequency_adjust * 0.8))
|
||||
elif "过少" in response:
|
||||
logger.info(f"频率调整: 过少，调整值到{final_value_by_api}")
|
||||
self.talk_frequency_adjust = max(0.1, min(3.0, self.talk_frequency_adjust * 1.2))
|
||||
self.talk_frequency_adjust = max(0.1, min(1.5, self.talk_frequency_adjust * 1.2))
|
||||
self.last_frequency_adjust_time = time.time()
|
||||
else:
|
||||
logger.info(f"频率调整：response不符合要求，取消本次调整")
|
||||
|
|
|
|||
|
|
@ -690,9 +690,6 @@ class StatisticOutputTask(AsyncTask):
|
|||
online_hours = stats[ONLINE_TIME] / 3600.0 if stats[ONLINE_TIME] > 0 else 0.0
|
||||
cost_per_hour = stats[TOTAL_COST] / online_hours if online_hours > 0 else 0.0
|
||||
|
||||
# 计算token/消息数量指标（每100条）
|
||||
tokens_per_100_messages = (total_tokens / stats[TOTAL_MSG_CNT] * 100) if stats[TOTAL_MSG_CNT] > 0 else 0.0
|
||||
|
||||
# 计算token/时间指标（token/小时）
|
||||
tokens_per_hour = (total_tokens / online_hours) if online_hours > 0 else 0.0
|
||||
|
||||
|
|
@ -700,8 +697,9 @@ class StatisticOutputTask(AsyncTask):
|
|||
total_replies = stats.get(TOTAL_REPLY_CNT, 0)
|
||||
cost_per_100_replies = (stats[TOTAL_COST] / total_replies * 100) if total_replies > 0 else 0.0
|
||||
|
||||
# 计算token/回复数量指标（每100条）
|
||||
tokens_per_100_replies = (total_tokens / total_replies * 100) if total_replies > 0 else 0.0
|
||||
# 计算花费/消息数量（排除自己回复）指标（每100条）
|
||||
total_messages_excluding_replies = stats[TOTAL_MSG_CNT] - total_replies
|
||||
cost_per_100_messages_excluding_replies = (stats[TOTAL_COST] / total_messages_excluding_replies * 100) if total_messages_excluding_replies > 0 else 0.0
|
||||
|
||||
output = [
|
||||
f"总在线时间: {_format_online_time(stats[ONLINE_TIME])}",
|
||||
|
|
@ -711,10 +709,9 @@ class StatisticOutputTask(AsyncTask):
|
|||
f"总Token数: {_format_large_number(total_tokens)}",
|
||||
f"总花费: {stats[TOTAL_COST]:.2f}¥",
|
||||
f"花费/消息数量: {cost_per_100_messages:.4f}¥/100条" if stats[TOTAL_MSG_CNT] > 0 else "花费/消息数量: N/A",
|
||||
f"花费/回复数量: {cost_per_100_replies:.4f}¥/100条" if total_replies > 0 else "花费/回复数量: N/A",
|
||||
f"花费/接受消息数量: {cost_per_100_messages_excluding_replies:.4f}¥/100条" if total_messages_excluding_replies > 0 else "花费/消息数量(排除回复): N/A",
|
||||
f"花费/回复消息数量: {cost_per_100_replies:.4f}¥/100条" if total_replies > 0 else "花费/回复数量: N/A",
|
||||
f"花费/时间: {cost_per_hour:.2f}¥/小时" if online_hours > 0 else "花费/时间: N/A",
|
||||
f"Token/消息数量: {_format_large_number(tokens_per_100_messages)}/100条" if stats[TOTAL_MSG_CNT] > 0 else "Token/消息数量: N/A",
|
||||
f"Token/回复数量: {_format_large_number(tokens_per_100_replies)}/100条" if total_replies > 0 else "Token/回复数量: N/A",
|
||||
f"Token/时间: {_format_large_number(tokens_per_hour)}/小时" if online_hours > 0 else "Token/时间: N/A",
|
||||
"",
|
||||
]
|
||||
|
|
@ -933,25 +930,21 @@ class StatisticOutputTask(AsyncTask):
|
|||
<div class=\"kpi-value\">{(stat_data[TOTAL_COST] / stat_data[TOTAL_MSG_CNT] * 100 if stat_data[TOTAL_MSG_CNT] > 0 else 0.0):.4f} ¥/100条</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">花费/时间</div>
|
||||
<div class=\"kpi-value\">{(stat_data[TOTAL_COST] / (stat_data[ONLINE_TIME] / 3600.0) if stat_data[ONLINE_TIME] > 0 else 0.0):.2f} ¥/小时</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">Token/消息数量</div>
|
||||
<div class=\"kpi-value\">{_format_large_number(sum(stat_data[TOTAL_TOK_BY_MODEL].values()) / stat_data[TOTAL_MSG_CNT] * 100 if stat_data[TOTAL_MSG_CNT] > 0 and stat_data[TOTAL_TOK_BY_MODEL] else 0.0, html=True)}/100条</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">Token/回复数量</div>
|
||||
<div class=\"kpi-value\">{_format_large_number(sum(stat_data[TOTAL_TOK_BY_MODEL].values()) / stat_data.get(TOTAL_REPLY_CNT, 0) * 100 if stat_data.get(TOTAL_REPLY_CNT, 0) > 0 and stat_data[TOTAL_TOK_BY_MODEL] else 0.0, html=True)}/100条</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">Token/时间</div>
|
||||
<div class=\"kpi-value\">{_format_large_number(sum(stat_data[TOTAL_TOK_BY_MODEL].values()) / (stat_data[ONLINE_TIME] / 3600.0) if stat_data[ONLINE_TIME] > 0 and stat_data[TOTAL_TOK_BY_MODEL] else 0.0, html=True)}/小时</div>
|
||||
<div class=\"kpi-title\">花费/消息数量(排除回复)</div>
|
||||
<div class=\"kpi-value\">{(stat_data[TOTAL_COST] / (stat_data[TOTAL_MSG_CNT] - stat_data.get(TOTAL_REPLY_CNT, 0)) * 100 if (stat_data[TOTAL_MSG_CNT] - stat_data.get(TOTAL_REPLY_CNT, 0)) > 0 else 0.0):.4f} ¥/100条</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">花费/回复数量</div>
|
||||
<div class=\"kpi-value\">{(stat_data[TOTAL_COST] / stat_data.get(TOTAL_REPLY_CNT, 0) * 100 if stat_data.get(TOTAL_REPLY_CNT, 0) > 0 else 0.0):.4f} ¥/100条</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">花费/时间</div>
|
||||
<div class=\"kpi-value\">{(stat_data[TOTAL_COST] / (stat_data[ONLINE_TIME] / 3600.0) if stat_data[ONLINE_TIME] > 0 else 0.0):.2f} ¥/小时</div>
|
||||
</div>
|
||||
<div class=\"kpi-card\">
|
||||
<div class=\"kpi-title\">Token/时间</div>
|
||||
<div class=\"kpi-value\">{_format_large_number(sum(stat_data[TOTAL_TOK_BY_MODEL].values()) / (stat_data[ONLINE_TIME] / 3600.0) if stat_data[ONLINE_TIME] > 0 and stat_data[TOTAL_TOK_BY_MODEL] else 0.0, html=True)}/小时</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<h2>按模型分类统计</h2>
|
||||
|
|
@ -1805,10 +1798,8 @@ class StatisticOutputTask(AsyncTask):
|
|||
# 初始化数据结构
|
||||
cost_per_100_messages = [0.0] * len(time_points) # 花费/消息数量（每100条）
|
||||
cost_per_hour = [0.0] * len(time_points) # 花费/时间（每小时）
|
||||
tokens_per_100_messages = [0.0] * len(time_points) # Token/消息数量（每100条）
|
||||
tokens_per_hour = [0.0] * len(time_points) # Token/时间（每小时）
|
||||
cost_per_100_replies = [0.0] * len(time_points) # 花费/回复数量（每100条）
|
||||
tokens_per_100_replies = [0.0] * len(time_points) # Token/回复数量（每100条）
|
||||
|
||||
# 每个时间点的累计数据
|
||||
total_costs = [0.0] * len(time_points)
|
||||
|
|
@ -1882,10 +1873,6 @@ class StatisticOutputTask(AsyncTask):
|
|||
if total_online_hours[idx] > 0:
|
||||
cost_per_hour[idx] = (total_costs[idx] / total_online_hours[idx])
|
||||
|
||||
# Token/消息数量（每100条）
|
||||
if total_messages[idx] > 0:
|
||||
tokens_per_100_messages[idx] = (total_tokens[idx] / total_messages[idx] * 100)
|
||||
|
||||
# Token/时间（每小时）
|
||||
if total_online_hours[idx] > 0:
|
||||
tokens_per_hour[idx] = (total_tokens[idx] / total_online_hours[idx])
|
||||
|
|
@ -1893,10 +1880,6 @@ class StatisticOutputTask(AsyncTask):
|
|||
# 花费/回复数量（每100条）
|
||||
if total_replies[idx] > 0:
|
||||
cost_per_100_replies[idx] = (total_costs[idx] / total_replies[idx] * 100)
|
||||
|
||||
# Token/回复数量（每100条）
|
||||
if total_replies[idx] > 0:
|
||||
tokens_per_100_replies[idx] = (total_tokens[idx] / total_replies[idx] * 100)
|
||||
|
||||
# 生成时间标签
|
||||
if interval_hours == 1:
|
||||
|
|
@ -1908,10 +1891,8 @@ class StatisticOutputTask(AsyncTask):
|
|||
"time_labels": time_labels,
|
||||
"cost_per_100_messages": cost_per_100_messages,
|
||||
"cost_per_hour": cost_per_hour,
|
||||
"tokens_per_100_messages": tokens_per_100_messages,
|
||||
"tokens_per_hour": tokens_per_hour,
|
||||
"cost_per_100_replies": cost_per_100_replies,
|
||||
"tokens_per_100_replies": tokens_per_100_replies,
|
||||
}
|
||||
|
||||
def _generate_metrics_tab(self, metrics_data: dict) -> str:
|
||||
|
|
@ -1919,10 +1900,8 @@ class StatisticOutputTask(AsyncTask):
|
|||
colors = {
|
||||
"cost_per_100_messages": "#8b5cf6",
|
||||
"cost_per_hour": "#9f8efb",
|
||||
"tokens_per_100_messages": "#b5a6ff",
|
||||
"tokens_per_hour": "#c7bbff",
|
||||
"cost_per_100_replies": "#d9ceff",
|
||||
"tokens_per_100_replies": "#a78bfa",
|
||||
}
|
||||
|
||||
return f"""
|
||||
|
|
@ -1944,17 +1923,11 @@ class StatisticOutputTask(AsyncTask):
|
|||
<div style="margin-bottom: 40px;">
|
||||
<canvas id="costPerHourChart" width="800" height="400"></canvas>
|
||||
</div>
|
||||
<div style="margin-bottom: 40px;">
|
||||
<canvas id="tokensPer100MessagesChart" width="800" height="400"></canvas>
|
||||
</div>
|
||||
<div style="margin-bottom: 40px;">
|
||||
<canvas id="tokensPerHourChart" width="800" height="400"></canvas>
|
||||
</div>
|
||||
<div style="margin-bottom: 40px;">
|
||||
<canvas id="costPer100RepliesChart" width="800" height="400"></canvas>
|
||||
</div>
|
||||
<div>
|
||||
<canvas id="tokensPer100RepliesChart" width="800" height="400"></canvas>
|
||||
<canvas id="costPer100RepliesChart" width="800" height="400"></canvas>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
|
|
@ -2001,13 +1974,6 @@ class StatisticOutputTask(AsyncTask):
|
|||
dataKey: 'cost_per_hour',
|
||||
color: '{colors["cost_per_hour"]}'
|
||||
}},
|
||||
tokensPer100Messages: {{
|
||||
id: 'tokensPer100MessagesChart',
|
||||
title: 'Token/消息数量',
|
||||
yAxisLabel: 'Token (/100条)',
|
||||
dataKey: 'tokens_per_100_messages',
|
||||
color: '{colors["tokens_per_100_messages"]}'
|
||||
}},
|
||||
tokensPerHour: {{
|
||||
id: 'tokensPerHourChart',
|
||||
title: 'Token/时间',
|
||||
|
|
@ -2021,13 +1987,6 @@ class StatisticOutputTask(AsyncTask):
|
|||
yAxisLabel: '花费 (¥/100条)',
|
||||
dataKey: 'cost_per_100_replies',
|
||||
color: '{colors["cost_per_100_replies"]}'
|
||||
}},
|
||||
tokensPer100Replies: {{
|
||||
id: 'tokensPer100RepliesChart',
|
||||
title: 'Token/回复数量',
|
||||
yAxisLabel: 'Token (/100条)',
|
||||
dataKey: 'tokens_per_100_replies',
|
||||
color: '{colors["tokens_per_100_replies"]}'
|
||||
}}
|
||||
}};
|
||||
|
||||
|
|
@ -2054,10 +2013,8 @@ class StatisticOutputTask(AsyncTask):
|
|||
// 重新创建图表
|
||||
createMetricsChart('costPer100Messages', data, timeScale);
|
||||
createMetricsChart('costPerHour', data, timeScale);
|
||||
createMetricsChart('tokensPer100Messages', data, timeScale);
|
||||
createMetricsChart('tokensPerHour', data, timeScale);
|
||||
createMetricsChart('costPer100Replies', data, timeScale);
|
||||
createMetricsChart('tokensPer100Replies', data, timeScale);
|
||||
}}
|
||||
|
||||
function createMetricsChart(chartType, data, timeScale) {{
|
||||
|
|
|
|||
|
|
@ -23,7 +23,7 @@ def init_memory_retrieval_prompt():
|
|||
# 第一步：问题生成prompt
|
||||
Prompt(
|
||||
"""
|
||||
你是一个专门检测是否需要回忆的助手。你的名字是{bot_name}。现在是{time_now}。
|
||||
你的名字是{bot_name}。现在是{time_now}。
|
||||
群里正在进行的聊天内容：
|
||||
{chat_history}
|
||||
|
||||
|
|
@ -34,6 +34,7 @@ def init_memory_retrieval_prompt():
|
|||
1. 对话中是否提到了过去发生的事情、人物、事件或信息
|
||||
2. 是否有需要回忆的内容（比如"之前说过"、"上次"、"以前"等）
|
||||
3. 是否有需要查找历史信息的问题
|
||||
4. 是否有问题可以搜集信息帮助你聊天
|
||||
|
||||
重要提示：
|
||||
- 如果"最近已查询的问题和结果"中已经包含了类似的问题，请避免重复生成相同或相似的问题
|
||||
|
|
@ -65,7 +66,9 @@ def init_memory_retrieval_prompt():
|
|||
|
||||
# 第二步：ReAct Agent prompt（工具描述会在运行时动态生成）
|
||||
Prompt(
|
||||
"""你需要通过思考(Think)、行动(Action)、观察(Observation)的循环来回答问题。
|
||||
"""
|
||||
你的名字是{bot_name}，你正在参与聊天，你需要搜集信息来回答问题，帮助你参与聊天。
|
||||
你需要通过思考(Think)、行动(Action)、观察(Observation)的循环来回答问题。
|
||||
|
||||
当前问题：{question}
|
||||
已收集的信息：
|
||||
|
|
@ -78,14 +81,20 @@ def init_memory_retrieval_prompt():
|
|||
```json
|
||||
{{
|
||||
"thought": "你的思考过程，分析当前情况，决定下一步行动",
|
||||
"action_type": {action_types_list},
|
||||
"action_params": {{参数名: 参数值}} 或 null
|
||||
"actions": [
|
||||
{{
|
||||
"action_type": {action_types_list},
|
||||
"action_params": {{参数名: 参数值}} 或 null
|
||||
}}
|
||||
]
|
||||
}}
|
||||
```
|
||||
|
||||
你可以选择以下动作：
|
||||
1. 如果已经收集到足够的信息可以回答问题，请设置action_type为"final_answer"，并在thought中说明答案。除非明确找到答案，否则不要设置为final_answer。
|
||||
2. 如果经过多次查询后，确认无法找到相关信息或答案，请设置action_type为"no_answer"，并在thought中说明原因。
|
||||
重要说明：
|
||||
- 你可以在一次迭代中执行多个查询，将多个action放在actions数组中
|
||||
- 如果只需要执行一个查询，actions数组中只包含一个action即可
|
||||
- 如果已经收集到足够的信息可以回答问题，请设置actions为包含一个action_type为"final_answer"的数组，并在thought中说明答案。除非明确找到答案，否则不要设置为final_answer。
|
||||
- 如果经过多次查询后，确认无法找到相关信息或答案，请设置actions为包含一个action_type为"no_answer"的数组，并在thought中说明原因。
|
||||
|
||||
请只输出JSON，不要输出其他内容：
|
||||
""",
|
||||
|
|
@ -101,6 +110,8 @@ def _parse_react_response(response: str) -> Optional[Dict[str, Any]]:
|
|||
|
||||
Returns:
|
||||
Dict[str, Any]: 解析后的动作信息，如果解析失败返回None
|
||||
格式: {"thought": str, "actions": List[Dict[str, Any]]}
|
||||
每个action格式: {"action_type": str, "action_params": dict}
|
||||
"""
|
||||
try:
|
||||
# 尝试提取JSON（可能包含在```json代码块中）
|
||||
|
|
@ -123,6 +134,20 @@ def _parse_react_response(response: str) -> Optional[Dict[str, Any]]:
|
|||
logger.warning(f"解析的JSON不是对象格式: {action_info}")
|
||||
return None
|
||||
|
||||
# 确保actions字段存在且为列表
|
||||
if "actions" not in action_info:
|
||||
logger.warning(f"响应中缺少actions字段: {action_info}")
|
||||
return None
|
||||
|
||||
if not isinstance(action_info["actions"], list):
|
||||
logger.warning(f"actions字段不是数组格式: {action_info['actions']}")
|
||||
return None
|
||||
|
||||
# 确保actions不为空
|
||||
if len(action_info["actions"]) == 0:
|
||||
logger.warning("actions数组为空")
|
||||
return None
|
||||
|
||||
return action_info
|
||||
|
||||
except Exception as e:
|
||||
|
|
@ -163,9 +188,13 @@ async def _react_agent_solve_question(
|
|||
# 获取工具注册器
|
||||
tool_registry = get_tool_registry()
|
||||
|
||||
# 获取bot_name
|
||||
bot_name = global_config.bot.nickname
|
||||
|
||||
# 构建prompt（动态生成工具描述）
|
||||
prompt = await global_prompt_manager.format_prompt(
|
||||
"memory_retrieval_react_prompt",
|
||||
bot_name=bot_name,
|
||||
question=question,
|
||||
collected_info=collected_info if collected_info else "暂无信息",
|
||||
tools_description=tool_registry.get_tools_description(),
|
||||
|
|
@ -196,45 +225,50 @@ async def _react_agent_solve_question(
|
|||
break
|
||||
|
||||
thought = action_info.get("thought", "")
|
||||
action_type = action_info.get("action_type", "")
|
||||
action_params = action_info.get("action_params", {})
|
||||
actions = action_info.get("actions", [])
|
||||
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 思考: {thought}")
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 动作类型: {action_type}")
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 动作参数: {action_params}")
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 动作数量: {len(actions)}")
|
||||
|
||||
# 记录思考步骤
|
||||
# 记录思考步骤（包含所有actions）
|
||||
step = {
|
||||
"iteration": iteration + 1,
|
||||
"thought": thought,
|
||||
"action_type": action_type,
|
||||
"action_params": action_params,
|
||||
"observation": ""
|
||||
"actions": actions,
|
||||
"observations": []
|
||||
}
|
||||
|
||||
# 执行动作
|
||||
if action_type == "final_answer":
|
||||
# Agent认为已经找到答案
|
||||
answer = thought # 使用thought作为答案
|
||||
step["observation"] = "找到答案"
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 找到最终答案: {answer}")
|
||||
return True, answer, thinking_steps
|
||||
# 检查是否有final_answer或no_answer
|
||||
for action in actions:
|
||||
action_type = action.get("action_type", "")
|
||||
if action_type == "final_answer":
|
||||
# Agent认为已经找到答案
|
||||
answer = thought # 使用thought作为答案
|
||||
step["observations"] = ["找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 找到最终答案: {answer}")
|
||||
return True, answer, thinking_steps
|
||||
elif action_type == "no_answer":
|
||||
# Agent确认无法找到答案
|
||||
answer = thought # 使用thought说明无法找到答案的原因
|
||||
step["observations"] = ["确认无法找到答案"]
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 确认无法找到答案: {answer}")
|
||||
return False, answer, thinking_steps
|
||||
|
||||
elif action_type == "no_answer":
|
||||
# Agent确认无法找到答案
|
||||
answer = thought # 使用thought说明无法找到答案的原因
|
||||
step["observation"] = "确认无法找到答案"
|
||||
thinking_steps.append(step)
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 确认无法找到答案: {answer}")
|
||||
return False, answer, thinking_steps
|
||||
|
||||
# 使用工具注册器执行工具
|
||||
# 并行执行所有工具
|
||||
tool_registry = get_tool_registry()
|
||||
tool = tool_registry.get_tool(action_type)
|
||||
tool_tasks = []
|
||||
|
||||
if tool:
|
||||
try:
|
||||
for i, action in enumerate(actions):
|
||||
action_type = action.get("action_type", "")
|
||||
action_params = action.get("action_params", {})
|
||||
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 动作 {i+1}/{len(actions)}: {action_type}({action_params})")
|
||||
|
||||
tool = tool_registry.get_tool(action_type)
|
||||
|
||||
if tool:
|
||||
# 准备工具参数（需要添加chat_id如果工具需要）
|
||||
tool_params = action_params.copy()
|
||||
|
||||
|
|
@ -244,31 +278,38 @@ async def _react_agent_solve_question(
|
|||
if "chat_id" in sig.parameters:
|
||||
tool_params["chat_id"] = chat_id
|
||||
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 执行工具: {action_type}({tool_params})")
|
||||
# 创建异步任务
|
||||
async def execute_single_tool(tool_instance, params, act_type, act_params, iter_num):
|
||||
try:
|
||||
observation = await tool_instance.execute(**params)
|
||||
param_str = ", ".join([f"{k}={v}" for k, v in act_params.items()])
|
||||
return f"查询{act_type}({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 动作 {act_type} {error_msg}")
|
||||
return f"查询{act_type}失败: {error_msg}"
|
||||
|
||||
# 执行工具
|
||||
observation = await tool.execute(**tool_params)
|
||||
step["observation"] = observation
|
||||
tool_tasks.append(execute_single_tool(tool, tool_params, action_type, action_params, iteration))
|
||||
else:
|
||||
error_msg = f"未知的工具类型: {action_type}"
|
||||
logger.warning(f"ReAct Agent 第 {iteration + 1} 次迭代 动作 {i+1}/{len(actions)} {error_msg}")
|
||||
tool_tasks.append(asyncio.create_task(asyncio.sleep(0, result=f"查询{action_type}失败: {error_msg}")))
|
||||
|
||||
# 并行执行所有工具
|
||||
if tool_tasks:
|
||||
observations = await asyncio.gather(*tool_tasks, return_exceptions=True)
|
||||
|
||||
# 处理执行结果
|
||||
for i, observation in enumerate(observations):
|
||||
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 动作 {i+1} 执行异常: {observation}")
|
||||
|
||||
# 构建收集信息的描述
|
||||
param_str = ", ".join([f"{k}={v}" for k, v in action_params.items()])
|
||||
collected_info += f"\n查询{action_type}({param_str})的结果：{observation}\n"
|
||||
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 工具执行结果: {observation}")
|
||||
except Exception as e:
|
||||
error_msg = f"工具执行失败: {str(e)}"
|
||||
step["observation"] = error_msg
|
||||
logger.error(f"ReAct Agent 第 {iteration + 1} 次迭代 {error_msg}")
|
||||
else:
|
||||
step["observation"] = f"未知的工具类型: {action_type}"
|
||||
logger.warning(f"ReAct Agent 第 {iteration + 1} 次迭代 未知的工具类型: {action_type}")
|
||||
step["observations"].append(observation)
|
||||
collected_info += f"\n{observation}\n"
|
||||
logger.info(f"ReAct Agent 第 {iteration + 1} 次迭代 动作 {i+1} 执行结果: {observation}")
|
||||
|
||||
thinking_steps.append(step)
|
||||
|
||||
# 如果观察结果为空或无效，继续下一轮
|
||||
if step["observation"] and "无有效信息" not in step["observation"] and "未找到" not in step["observation"]:
|
||||
# 有有效信息，继续思考
|
||||
pass
|
||||
|
||||
# 达到最大迭代次数或超时，但Agent没有明确返回final_answer
|
||||
# 迭代超时应该直接视为no_answer，而不是使用已有信息
|
||||
|
|
@ -333,6 +374,47 @@ def _get_recent_query_history(chat_id: str, time_window_seconds: float = 300.0)
|
|||
return ""
|
||||
|
||||
|
||||
def _get_cached_memories(chat_id: str, time_window_seconds: float = 300.0) -> List[str]:
|
||||
"""获取最近一段时间内缓存的记忆（只返回找到答案的记录）
|
||||
|
||||
Args:
|
||||
chat_id: 聊天ID
|
||||
time_window_seconds: 时间窗口（秒），默认300秒（5分钟）
|
||||
|
||||
Returns:
|
||||
List[str]: 格式化的记忆列表，每个元素格式为 "问题：xxx\n答案：xxx"
|
||||
"""
|
||||
try:
|
||||
current_time = time.time()
|
||||
start_time = current_time - time_window_seconds
|
||||
|
||||
# 查询最近时间窗口内找到答案的记录，按更新时间倒序
|
||||
records = (
|
||||
ThinkingBack.select()
|
||||
.where(
|
||||
(ThinkingBack.chat_id == chat_id) &
|
||||
(ThinkingBack.update_time >= start_time) &
|
||||
(ThinkingBack.found_answer == 1)
|
||||
)
|
||||
.order_by(ThinkingBack.update_time.desc())
|
||||
.limit(5) # 最多返回5条最近的记录
|
||||
)
|
||||
|
||||
if not records.exists():
|
||||
return []
|
||||
|
||||
cached_memories = []
|
||||
for record in records:
|
||||
if record.answer:
|
||||
cached_memories.append(f"问题：{record.question}\n答案：{record.answer}")
|
||||
|
||||
return cached_memories
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"获取缓存记忆失败: {e}")
|
||||
return []
|
||||
|
||||
|
||||
def _query_thinking_back(chat_id: str, question: str) -> Optional[Tuple[bool, str]]:
|
||||
"""从thinking_back数据库中查询是否有现成的答案
|
||||
|
||||
|
|
@ -441,11 +523,11 @@ def _get_max_iterations_by_question_count(question_count: int) -> int:
|
|||
int: 最大迭代次数
|
||||
"""
|
||||
if question_count == 1:
|
||||
return 5
|
||||
return 6
|
||||
elif question_count == 2:
|
||||
return 3
|
||||
else: # 3个或以上
|
||||
return 1
|
||||
return 2
|
||||
|
||||
|
||||
async def _process_single_question(
|
||||
|
|
@ -587,9 +669,19 @@ async def build_memory_retrieval_prompt(
|
|||
# 解析问题列表
|
||||
questions = _parse_questions_json(response)
|
||||
|
||||
# 获取缓存的记忆（与question时使用相同的时间窗口和数量限制）
|
||||
cached_memories = _get_cached_memories(chat_id, time_window_seconds=300.0)
|
||||
|
||||
if not questions:
|
||||
logger.debug("模型认为不需要检索记忆或解析失败")
|
||||
return ""
|
||||
# 即使没有当次查询，也返回缓存的记忆
|
||||
if cached_memories:
|
||||
retrieved_memory = "\n\n".join(cached_memories)
|
||||
end_time = time.time()
|
||||
logger.info(f"无当次查询，返回缓存记忆，耗时: {(end_time - start_time):.3f}秒，包含 {len(cached_memories)} 条缓存记忆")
|
||||
return f"你回忆起了以下信息：\n{retrieved_memory}\n请在回复时参考这些回忆的信息。\n"
|
||||
else:
|
||||
return ""
|
||||
|
||||
logger.info(f"解析到 {len(questions)} 个问题: {questions}")
|
||||
|
||||
|
|
@ -613,20 +705,34 @@ async def build_memory_retrieval_prompt(
|
|||
|
||||
# 收集所有有效结果
|
||||
all_results = []
|
||||
current_questions = set() # 用于去重，避免缓存和当次查询重复
|
||||
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}")
|
||||
elif result is not None:
|
||||
all_results.append(result)
|
||||
# 提取问题用于去重
|
||||
if result.startswith("问题："):
|
||||
question = result.split("\n")[0].replace("问题：", "").strip()
|
||||
current_questions.add(question)
|
||||
|
||||
# 将缓存的记忆添加到结果中（排除当次查询已包含的问题，避免重复）
|
||||
for cached_memory in cached_memories:
|
||||
if cached_memory.startswith("问题："):
|
||||
question = cached_memory.split("\n")[0].replace("问题：", "").strip()
|
||||
# 只有当次查询中没有相同问题时，才添加缓存记忆
|
||||
if question not in current_questions:
|
||||
all_results.append(cached_memory)
|
||||
logger.debug(f"添加缓存记忆: {question[:50]}...")
|
||||
|
||||
end_time = time.time()
|
||||
|
||||
if all_results:
|
||||
retrieved_memory = "\n\n".join(all_results)
|
||||
logger.info(f"记忆检索成功，耗时: {(end_time - start_time):.3f}秒")
|
||||
logger.info(f"记忆检索成功，耗时: {(end_time - start_time):.3f}秒，包含 {len(all_results)} 条记忆（含缓存）")
|
||||
return f"你回忆起了以下信息：\n{retrieved_memory}\n请在回复时参考这些回忆的信息。\n"
|
||||
else:
|
||||
logger.debug("所有问题均未找到答案")
|
||||
logger.debug("所有问题均未找到答案，且无缓存记忆")
|
||||
return ""
|
||||
|
||||
except Exception as e:
|
||||
|
|
|
|||
Loading…
Reference in New Issue