fix:优化记忆提取和聊天压缩

This commit is contained in:
SengokuCola
2025-11-10 12:27:54 +08:00
parent 10cd2474af
commit 71a2a4282b
5 changed files with 212 additions and 139 deletions

View File

@@ -269,7 +269,16 @@ class ChatHistorySummarizer:
)
# 使用LLM压缩聊天内容
theme, keywords, summary = await self._compress_with_llm(original_text)
success, theme, keywords, summary = await self._compress_with_llm(original_text)
if not success:
logger.warning(
f"{self.log_prefix} LLM压缩失败不存储到数据库 | 消息数: {len(messages)}"
)
# 清空当前批次,避免重复处理
self.current_batch = None
return
logger.info(
f"{self.log_prefix} LLM压缩完成 | 主题: {theme} | 关键词数: {len(keywords)} | 概括长度: {len(summary)}"
)
@@ -297,12 +306,12 @@ class ChatHistorySummarizer:
# 出错时也清空批次,避免重复处理
self.current_batch = None
async def _compress_with_llm(self, original_text: str) -> tuple[str, List[str], str]:
async def _compress_with_llm(self, original_text: str) -> tuple[bool, str, List[str], str]:
"""
使用LLM压缩聊天内容
Returns:
tuple[str, List[str], str]: (主题, 关键词列表, 概括)
tuple[bool, str, List[str], str]: (是否成功, 主题, 关键词列表, 概括)
"""
prompt = f"""请对以下聊天记录进行概括,提取以下信息:
@@ -353,13 +362,13 @@ class ChatHistorySummarizer:
if isinstance(keywords, str):
keywords = [keywords]
return theme, keywords, summary
return True, theme, keywords, summary
except Exception as e:
logger.error(f"{self.log_prefix} LLM压缩聊天内容时出错: {e}")
logger.error(f"{self.log_prefix} LLM响应: {response if 'response' in locals() else 'N/A'}")
# 返回默认值
return "未命名对话", [], "压缩失败,无法生成概括"
# 返回失败标志和默认值
return False, "未命名对话", [], "压缩失败,无法生成概括"
async def _store_to_database(
self,

View File

@@ -221,14 +221,17 @@ def split_into_sentences_w_remove_punctuation(text: str) -> list[str]:
while i < len(text):
char = text[i]
if char in separators:
# 检查分割条件:如果空格左右都是英文字母,则不分割(仅对空格应用此规则)
# 检查分割条件:如果空格左右都是英文字母、数字,或数字和英文之间,则不分割(仅对空格应用此规则)
can_split = True
if 0 < i < len(text) - 1:
prev_char = text[i - 1]
next_char = text[i + 1]
# 只对空格应用"不分割两个英文之间的空格"规则
if char == ' ' and is_english_letter(prev_char) and is_english_letter(next_char):
can_split = False
# 只对空格应用"不分割数字和数字、数字和英文、英文和数字、英文和英文之间的空格"规则
if char == ' ':
prev_is_alnum = prev_char.isdigit() or is_english_letter(prev_char)
next_is_alnum = next_char.isdigit() or is_english_letter(next_char)
if prev_is_alnum and next_is_alnum:
can_split = False
if can_split:
# 只有当当前段不为空时才添加