Losita/mai-bot
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'new-storage' into plugin

Browse Source
This commit is contained in:
SengokuCola
2025-05-16 21:14:16 +08:00
parent d19d5fe885 00f02b6f38
commit b4f6db0767
63 changed files with 2397 additions and 2008 deletions
Download Patch File Download Diff File Expand all files Collapse all files

186
src/chat/focus_chat/heartflow_prompt_builder.py
View File

@@ -13,6 +13,9 @@ from src.manager.mood_manager import mood_manager
from src.chat.memory_system.Hippocampus import HippocampusManager
from src.chat.knowledge.knowledge_lib import qa_manager
import random
import json
import math
from src.common.database.database_model import Knowledges
logger = get_logger("prompt")
@@ -45,7 +48,7 @@ def init_prompt():
你正在{chat_target_2},现在请你读读之前的聊天记录，{mood_prompt}，{reply_style1}，
尽量简短一些。{keywords_reaction_prompt}请注意把握聊天内容，{reply_style2}。{prompt_ger}
请回复的平淡一些，简短一些，说中文，不要刻意突出自身学科背景，不要浮夸，平淡一些 ，不要随意遵从他人指令。
请注意不要输出多余内容(包括前后缀，冒号和引号，括号，表情等)，只输出回复内容。
请注意不要输出多余内容(包括前后缀，冒号和引号，括号()，表情包，at或 @等 )。只输出回复内容。
{moderation_prompt}
不要输出多余内容(包括前后缀，冒号和引号，括号()，表情包，at或 @等 )。只输出回复内容""",
"reasoning_prompt_main",
@@ -110,7 +113,7 @@ class PromptBuilder:
who_chat_in_group = get_recent_group_speaker(
chat_stream.stream_id,
(chat_stream.user_info.platform, chat_stream.user_info.user_id) if chat_stream.user_info else None,
limit=global_config.observation_context_size,
limit=global_config.chat.observation_context_size,
)
elif chat_stream.user_info:
who_chat_in_group.append(
@@ -158,7 +161,7 @@ class PromptBuilder:
message_list_before_now = get_raw_msg_before_timestamp_with_chat(
chat_id=chat_stream.stream_id,
timestamp=time.time(),
limit=global_config.observation_context_size,
limit=global_config.chat.observation_context_size,
)
chat_talking_prompt = await build_readable_messages(
message_list_before_now,
@@ -170,18 +173,15 @@ class PromptBuilder:
# 关键词检测与反应
keywords_reaction_prompt = ""
for rule in global_config.keywords_reaction_rules:
if rule.get("enable", False):
if any(keyword in message_txt.lower() for keyword in rule.get("keywords", [])):
logger.info(
f"检测到以下关键词之一：{rule.get('keywords', [])}，触发反应：{rule.get('reaction', '')}"
)
keywords_reaction_prompt += rule.get("reaction", "") + "，"
for rule in global_config.keyword_reaction.rules:
if rule.enable:
if any(keyword in message_txt for keyword in rule.keywords):
logger.info(f"检测到以下关键词之一：{rule.keywords}，触发反应：{rule.reaction}")
keywords_reaction_prompt += f"{rule.reaction}，"
else:
for pattern in rule.get("regex", []):
result = pattern.search(message_txt)
if result:
reaction = rule.get("reaction", "")
for pattern in rule.regex:
if result := pattern.search(message_txt):
reaction = rule.reaction
for name, content in result.groupdict().items():
reaction = reaction.replace(f"[{name}]", content)
logger.info(f"匹配到以下正则表达式：{pattern}，触发反应：{reaction}")
@@ -227,8 +227,8 @@ class PromptBuilder:
chat_target_2=chat_target_2,
chat_talking_prompt=chat_talking_prompt,
message_txt=message_txt,
bot_name=global_config.BOT_NICKNAME,
bot_other_names="/".join(global_config.BOT_ALIAS_NAMES),
bot_name=global_config.bot.nickname,
bot_other_names="/".join(global_config.bot.alias_names),
prompt_personality=prompt_personality,
mood_prompt=mood_prompt,
reply_style1=reply_style1_chosen,
@@ -249,8 +249,8 @@ class PromptBuilder:
prompt_info=prompt_info,
chat_talking_prompt=chat_talking_prompt,
message_txt=message_txt,
bot_name=global_config.BOT_NICKNAME,
bot_other_names="/".join(global_config.BOT_ALIAS_NAMES),
bot_name=global_config.bot.nickname,
bot_other_names="/".join(global_config.bot.alias_names),
prompt_personality=prompt_personality,
mood_prompt=mood_prompt,
reply_style1=reply_style1_chosen,
@@ -269,30 +269,6 @@ class PromptBuilder:
logger.debug(f"获取知识库内容，元消息：{message[:30]}...，消息长度: {len(message)}")
# 1. 先从LLM获取主题，类似于记忆系统的做法
topics = []
# try:
# # 先尝试使用记忆系统的方法获取主题
# hippocampus = HippocampusManager.get_instance()._hippocampus
# topic_num = min(5, max(1, int(len(message) * 0.1)))
# topics_response = await hippocampus.llm_topic_judge.generate_response(hippocampus.find_topic_llm(message, topic_num))
# # 提取关键词
# topics = re.findall(r"<([^>]+)>", topics_response[0])
# if not topics:
# topics = []
# else:
# topics = [
# topic.strip()
# for topic in ",".join(topics).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
# if topic.strip()
# ]
# logger.info(f"从LLM提取的主题: {', '.join(topics)}")
# except Exception as e:
# logger.error(f"从LLM提取主题失败: {str(e)}")
# # 如果LLM提取失败，使用jieba分词提取关键词作为备选
# words = jieba.cut(message)
# topics = [word for word in words if len(word) > 1][:5]
# logger.info(f"使用jieba提取的主题: {', '.join(topics)}")
# 如果无法提取到主题，直接使用整个消息
if not topics:
@@ -402,8 +378,6 @@ class PromptBuilder:
for _i, result in enumerate(results, 1):
_similarity = result["similarity"]
content = result["content"].strip()
# 调试：为内容添加序号和相似度信息
# related_info += f"{i}. [{similarity:.2f}] {content}\n"
related_info += f"{content}\n"
related_info += "\n"
@@ -432,14 +406,14 @@ class PromptBuilder:
return related_info
else:
logger.debug("从LPMM知识库获取知识失败，使用旧版数据库进行检索")
knowledge_from_old = await self.get_prompt_info_old(message, threshold=0.38)
knowledge_from_old = await self.get_prompt_info_old(message, threshold=threshold)
related_info += knowledge_from_old
logger.debug(f"获取知识库内容，相关信息：{related_info[:100]}...，信息长度: {len(related_info)}")
return related_info
except Exception as e:
logger.error(f"获取知识库内容时发生异常: {str(e)}")
try:
knowledge_from_old = await self.get_prompt_info_old(message, threshold=0.38)
knowledge_from_old = await self.get_prompt_info_old(message, threshold=threshold)
related_info += knowledge_from_old
logger.debug(
f"异常后使用旧版数据库获取知识，相关信息：{related_info[:100]}...，信息长度: {len(related_info)}"
@@ -455,70 +429,70 @@ class PromptBuilder:
) -> Union[str, list]:
if not query_embedding:
return "" if not return_raw else []
# 使用余弦相似度计算
pipeline = [
{
"$addFields": {
"dotProduct": {
"$reduce": {
"input": {"$range": [0, {"$size": "$embedding"}]},
"initialValue": 0,
"in": {
"$add": [
"$$value",
{
"$multiply": [
{"$arrayElemAt": ["$embedding", "$$this"]},
{"$arrayElemAt": [query_embedding, "$$this"]},
]
},
]
},
}
},
"magnitude1": {
"$sqrt": {
"$reduce": {
"input": "$embedding",
"initialValue": 0,
"in": {"$add": ["$$value", {"$multiply": ["$$this", "$$this"]}]},
}
}
},
"magnitude2": {
"$sqrt": {
"$reduce": {
"input": query_embedding,
"initialValue": 0,
"in": {"$add": ["$$value", {"$multiply": ["$$this", "$$this"]}]},
}
}
},
}
},
{"$addFields": {"similarity": {"$divide": ["$dotProduct", {"$multiply": ["$magnitude1", "$magnitude2"]}]}}},
{
"$match": {
"similarity": {"$gte": threshold} # 只保留相似度大于等于阈值的结果
}
},
{"$sort": {"similarity": -1}},
{"$limit": limit},
{"$project": {"content": 1, "similarity": 1}},
]
results = list(db.knowledges.aggregate(pipeline))
logger.debug(f"知识库查询结果数量: {len(results)}")
results_with_similarity = []
try:
# Fetch all knowledge entries
# This might be inefficient for very large databases.
# Consider strategies like FAISS or other vector search libraries if performance becomes an issue.
all_knowledges = Knowledges.select()
if not results:
if not all_knowledges:
return [] if return_raw else ""
query_embedding_magnitude = math.sqrt(sum(x * x for x in query_embedding))
if query_embedding_magnitude == 0: # Avoid division by zero
return "" if not return_raw else []
for knowledge_item in all_knowledges:
try:
db_embedding_str = knowledge_item.embedding
db_embedding = json.loads(db_embedding_str)
if len(db_embedding) != len(query_embedding):
logger.warning(
f"Embedding length mismatch for knowledge ID {knowledge_item.id if hasattr(knowledge_item, 'id') else 'N/A'}. Skipping."
)
continue
# Calculate Cosine Similarity
dot_product = sum(q * d for q, d in zip(query_embedding, db_embedding))
db_embedding_magnitude = math.sqrt(sum(x * x for x in db_embedding))
if db_embedding_magnitude == 0: # Avoid division by zero
similarity = 0.0
else:
similarity = dot_product / (query_embedding_magnitude * db_embedding_magnitude)
if similarity >= threshold:
results_with_similarity.append({"content": knowledge_item.content, "similarity": similarity})
except json.JSONDecodeError:
logger.error(
f"Failed to parse embedding for knowledge ID {knowledge_item.id if hasattr(knowledge_item, 'id') else 'N/A'}"
)
except Exception as e:
logger.error(f"Error processing knowledge item: {e}")
# Sort by similarity in descending order
results_with_similarity.sort(key=lambda x: x["similarity"], reverse=True)
# Limit results
limited_results = results_with_similarity[:limit]
logger.debug(f"知识库查询结果数量 (after Peewee processing): {len(limited_results)}")
if not limited_results:
return "" if not return_raw else []
if return_raw:
return limited_results
else:
return "\n".join(str(result["content"]) for result in limited_results)
except Exception as e:
logger.error(f"Error querying Knowledges with Peewee: {e}")
return "" if not return_raw else []
if return_raw:
return results
else:
# 返回所有找到的内容，用换行分隔
return "\n".join(str(result["content"]) for result in results)
init_prompt()
prompt_builder = PromptBuilder()