perf: 提升上下文压缩保真度，避免二次失真
- 摘要前单条消息截断上限 200→1500 字，且长内容优先保留代码块（编码场景不丢上下文） - keep_recent 硬编码 10→可配置（默认 12），适配跨多轮的编码编辑 - 历史摘要标记不再被二次摘要：累积合并为单条 system，新消息才产生新摘要块 - 摘要按角色分条、要求保留工具调用/代码修改点；simple_summarize 额外提取编辑文件 - 新增 tests/test_compress_quality.py 覆盖工具JSON保真/代码块保留/幂等/可配置keep_recent Co-Authored-By: WorkBuddy <workbuddy@tencent.com>
This commit is contained in:
shuiyu67andWorkBuddy committed
1 parent
5fe1a7c868
commit
e73744c742
2 files changed
+180
-29
No files matched your search
+77
-29
@@ -23,7 +23,10 @@ class MemoryManager:
|
||||
self.memory_file = os.path.join(project_dir, "MEMORY.md")
|
||||
self.chat_dir = os.path.join(project_dir, "chat_history")
|
||||
self.compress_threshold = 50 # 超过 50 条消息触发压缩（兜底）
|
||||
self.compress_ratio = 0.7 # token 感知压缩阈值：used/context_window >= 0.7 触发
|
||||
self.compress_ratio = 0.7 # token 感知压缩阈值：used/max_input_tokens >= 0.7 触发
|
||||
self.keep_recent = 12 # 压缩时保留最近 N 条（编码常跨多轮，略高于原 10）
|
||||
self.max_summary_chars = 1500 # 摘要前单条消息截断上限（原硬编码 200，编码场景丢代码）
|
||||
self.compressed_blobs = [] # 累积的压缩摘要（避免二次摘要失真）
|
||||
self._ensure_dirs()
|
||||
|
||||
def _ensure_dirs(self):
|
||||
@@ -324,52 +327,88 @@ class MemoryManager:
|
||||
pass
|
||||
return len(conversation) > self.compress_threshold
|
||||
|
||||
def compress_context(self, conversation: list, engine=None) -> list:
|
||||
def _is_compressed_marker(self, m: dict) -> bool:
|
||||
"""判断一条消息是否是我们自己产出的压缩摘要（避免二次摘要失真）"""
|
||||
return bool(m.get("role") == "system"
|
||||
and str(m.get("content", "")).startswith("[上下文压缩]"))
|
||||
|
||||
def _prepare_for_summary(self, msg: dict) -> str:
|
||||
"""把单条消息转成摘要输入：保真优先——保留工具调用 JSON，长内容截断但尽量留代码块。"""
|
||||
role = msg.get("role", "unknown")
|
||||
content = msg.get("content", "")
|
||||
if not isinstance(content, str):
|
||||
return f"[{role}]: (非文本/结构化内容，长度 {len(str(content))})\n"
|
||||
|
||||
is_tool = ('"action": "use_tool"' in content) or ('"tool":' in content)
|
||||
cap = self.max_summary_chars
|
||||
|
||||
if len(content) <= cap:
|
||||
body = content
|
||||
else:
|
||||
# 尽量保留代码块（```...```）以保真编码上下文
|
||||
fences = re.findall(r"```.*?```", content, re.DOTALL)
|
||||
if fences:
|
||||
kept = "\n".join(fences[:2])
|
||||
if len(kept) > cap:
|
||||
kept = kept[:cap] + "\n...(截断)"
|
||||
body = f"(内容过长已截断，保留代码块)\n{kept}"
|
||||
else:
|
||||
body = content[:cap] + "\n...(截断)"
|
||||
|
||||
marker = " [工具调用]" if is_tool else ""
|
||||
return f"[{role}]{marker}: {body}\n"
|
||||
|
||||
def compress_context(self, conversation: list, engine=None, keep_recent=None) -> list:
|
||||
"""
|
||||
压缩对话历史
|
||||
策略: 保留最近 10 条 + 用 AI 总结前面的内容
|
||||
压缩对话历史。
|
||||
策略: 保留最近 keep_recent 条 + 用当前引擎 AI 总结前面的内容。
|
||||
关键改进:
|
||||
- keep_recent 可配置（默认 12，编码常跨多轮）
|
||||
- 已压缩摘要(标记消息)不再二次摘要，避免失真累积
|
||||
- 摘要按角色分条，保留工具调用/代码修改结构
|
||||
"""
|
||||
if len(conversation) <= 10:
|
||||
if keep_recent is None:
|
||||
keep_recent = self.keep_recent
|
||||
if len(conversation) <= keep_recent:
|
||||
return conversation
|
||||
|
||||
keep_recent = 10
|
||||
to_compress = conversation[:-keep_recent]
|
||||
recent = conversation[-keep_recent:]
|
||||
prefix = conversation[:-keep_recent]
|
||||
# 只摘要尚未压缩过的"真消息"，跳过历史摘要本身 → 不二次失真
|
||||
new_to_compress = [m for m in prefix if not self._is_compressed_marker(m)]
|
||||
if not new_to_compress:
|
||||
# 前缀全是历史摘要，没有新内容可压，保持现状即可（不丢摘要）
|
||||
return conversation
|
||||
|
||||
# 尝试用 AI 生成摘要
|
||||
summary = self._ai_summarize(to_compress, engine)
|
||||
summary = self._ai_summarize(new_to_compress, engine)
|
||||
block = f"[上下文压缩] 之前 {len(new_to_compress)} 条对话摘要:\n{summary}"
|
||||
self.compressed_blobs.append(block)
|
||||
|
||||
# 构建压缩后的历史
|
||||
compressed = [{
|
||||
"role": "system",
|
||||
"content": f"[上下文压缩] 以下是之前 {len(to_compress)} 条对话的摘要:\n{summary}"
|
||||
}]
|
||||
# 累积所有摘要合并为单条 system（而非每次重压旧摘要）
|
||||
merged = "\n\n".join(self.compressed_blobs)
|
||||
compressed = [{"role": "system", "content": merged}]
|
||||
compressed.extend(recent)
|
||||
return compressed
|
||||
|
||||
def _ai_summarize(self, messages: list, engine=None) -> str:
|
||||
"""用 AI 生成对话摘要"""
|
||||
"""用当前引擎生成对话摘要（保真优先）"""
|
||||
if not engine:
|
||||
return self._simple_summarize(messages)
|
||||
|
||||
# 构建摘要请求
|
||||
conversation_text = ""
|
||||
for msg in messages:
|
||||
role = msg.get("role", "unknown")
|
||||
content = msg.get("content", "")
|
||||
if isinstance(content, str):
|
||||
conversation_text += f"[{role}]: {content[:200]}\n"
|
||||
conversation_text = "".join(self._prepare_for_summary(m) for m in messages)
|
||||
|
||||
prompt = (
|
||||
"请用简洁的中文总结以下对话的关键信息（工具调用、代码修改、重要决策）。"
|
||||
"控制在 200 字以内。\n\n"
|
||||
"请用简洁中文总结以下对话的关键信息，要求:\n"
|
||||
"1. 按角色(用户/助手)分条，保留每条的要点；\n"
|
||||
"2. 重点保留: 工具调用名称与参数、代码修改点(读/改/创建的文件与位置)、重要决策；\n"
|
||||
"3. 控制总体在 400 字以内，不要编造未提及的信息。\n\n"
|
||||
f"{conversation_text}"
|
||||
)
|
||||
|
||||
try:
|
||||
summary = engine.generate_response(
|
||||
prompt,
|
||||
system_prompt="你是摘要生成器。只输出摘要，不要多余内容。"
|
||||
system_prompt="你是摘要生成器。只输出结构化摘要，不要多余内容。"
|
||||
)
|
||||
if summary and len(summary) > 10:
|
||||
return summary
|
||||
@@ -379,23 +418,25 @@ class MemoryManager:
|
||||
return self._simple_summarize(messages)
|
||||
|
||||
def _simple_summarize(self, messages: list) -> str:
|
||||
"""简单摘要（不依赖 AI）"""
|
||||
"""简单摘要（不依赖 AI，保真优先）"""
|
||||
topics = []
|
||||
tool_calls = []
|
||||
edits = []
|
||||
for msg in messages:
|
||||
content = msg.get("content", "")
|
||||
role = msg.get("role", "")
|
||||
if role == "user" and isinstance(content, str):
|
||||
if not isinstance(content, str):
|
||||
continue
|
||||
if role == "user":
|
||||
# 提取用户请求
|
||||
if len(content) < 100:
|
||||
topics.append(content)
|
||||
else:
|
||||
topics.append(content[:50] + "...")
|
||||
if role == "assistant" and isinstance(content, str):
|
||||
if role == "assistant":
|
||||
# 提取工具调用
|
||||
if '"action": "use_tool"' in content or '"tool":' in content:
|
||||
try:
|
||||
# 尝试解析工具名
|
||||
for part in content.split('"tool":'):
|
||||
if len(part) > 2:
|
||||
tool_name = part.split('"')[1] if '"' in part else ""
|
||||
@@ -403,12 +444,19 @@ class MemoryManager:
|
||||
tool_calls.append(tool_name)
|
||||
except Exception:
|
||||
pass
|
||||
# 提取代码编辑点（edit 命令）
|
||||
for m in re.finditer(r'edit\s+(\S+)', content):
|
||||
fn = m.group(1)
|
||||
if fn not in edits:
|
||||
edits.append(fn)
|
||||
|
||||
summary_parts = []
|
||||
if topics:
|
||||
summary_parts.append(f"讨论了 {len(topics)} 个话题: {'; '.join(topics[:5])}")
|
||||
if tool_calls:
|
||||
summary_parts.append(f"使用了工具: {', '.join(tool_calls[:8])}")
|
||||
if edits:
|
||||
summary_parts.append(f"编辑了文件: {', '.join(edits[:8])}")
|
||||
summary_parts.append(f"共 {len(messages)} 条对话")
|
||||
|
||||
return "。".join(summary_parts)
|
||||
|
||||
Reference in new issue
Block a user