process_message只保留find、get工具

This commit is contained in:
朱潮 2025-12-10 18:06:55 +08:00
parent 93d13f3d22
commit 4a909cd846

View File

@ -144,7 +144,15 @@ def process_messages(messages: List[Dict], language: Optional[str] = None) -> Li
这是 get_content_from_messages 的逆运算将包含 [TOOL_RESPONSE] 的消息重新组装回
msg['role'] == 'function' msg.get('function_call') 的格式
Args:
messages: 消息列表
language: 可选的语言参数
include_function_name: 需要包含的function_name关键词列表默认包含['find', 'get']
"""
# 设置默认的排除function_name列表
include_function_name = ['find', 'get']
processed_messages = []
# 收集所有ASSISTANT消息的索引
@ -173,38 +181,36 @@ def process_messages(messages: List[Dict], language: Optional[str] = None) -> Li
continue
# 不往后传输 历史工具调用的文字
if current_tag == "ANSWER":
filtered_content += f"[ANSWER]\n{text}\n"
# if current_tag == "TOOL_RESPONSE":
# if is_recent_message:
# # 最近10条ASSISTANT消息保留完整TOOL_RESPONSE信息使用简略模式
# if len(text) <= 500:
# filtered_content += f"[TOOL_RESPONSE]\n{text}\n"
# else:
# # 截取前中后3段内容每段250字
# first_part = text[:250]
# middle_start = len(text) // 2 - 125
# middle_part = text[middle_start:middle_start + 250]
# last_part = text[-250:]
if current_tag == "TOOL_RESPONSE":
if is_recent_message:
# 最近10条ASSISTANT消息保留完整TOOL_RESPONSE信息使用简略模式
if len(text) <= 1000:
filtered_content += f"[TOOL_RESPONSE] {text}\n"
else:
# 截取前中后3段内容每段250字
first_part = text[:250]
middle_start = len(text) // 2 - 125
middle_part = text[middle_start:middle_start + 250]
last_part = text[-250:]
# # 计算省略的字数
# omitted_count = len(text) - 750
# omitted_text = f"...此处省略{omitted_count}字..."
# 计算省略的字数
omitted_count = len(text) - 750
omitted_text = f"...此处省略{omitted_count}字..."
# # 拼接内容
# truncated_text = f"{first_part}\n{omitted_text}\n{middle_part}\n{omitted_text}\n{last_part}"
# filtered_content += f"[TOOL_RESPONSE]\n{truncated_text}\n"
# # 10条以上的消息不保留TOOL_RESPONSE数据完全跳过
# elif current_tag == "TOOL_CALL":
# if is_recent_message:
# # 最近10条ASSISTANT消息保留TOOL_CALL信息
# filtered_content += f"[TOOL_CALL]\n{text}\n"
# # 10条以上的消息不保留TOOL_CALL数据完全跳过
# elif current_tag == "ANSWER":
# # 所有ASSISTANT消息都保留ANSWER数据
# filtered_content += f"[ANSWER]\n{text}\n"
# elif current_tag != "THINK" and current_tag != "PREAMBLE":
# filtered_content += text + "\n"
# 拼接内容
truncated_text = f"{first_part}\n{omitted_text}\n{middle_part}\n{omitted_text}\n{last_part}"
filtered_content += f"[TOOL_RESPONSE] {truncated_text}\n"
# 10条以上的消息不保留TOOL_RESPONSE数据完全跳过
elif current_tag == "TOOL_CALL":
if is_recent_message:
# 最近10条ASSISTANT消息保留TOOL_CALL信息
filtered_content += f"[TOOL_CALL] {text}\n"
# 10条以上的消息不保留TOOL_CALL数据完全跳过
elif current_tag == "ANSWER":
# 所有ASSISTANT消息都保留ANSWER数据
filtered_content += f"[ANSWER] {text}\n"
elif current_tag != "THINK" and current_tag != "PREAMBLE":
filtered_content += text + "\n"
else: # 标签
current_tag = parts[i]
@ -237,48 +243,61 @@ def process_messages(messages: List[Dict], language: Optional[str] = None) -> Li
if not text:
continue
# 不往后传输 历史工具调用的文字
if current_tag == "ANSWER":
if current_tag == "TOOL_RESPONSE":
# 解析 TOOL_RESPONSE 格式:[TOOL_RESPONSE] function_name\ncontent
lines = text.split('\n', 1)
function_name = lines[0].strip() if lines else ""
response_content = lines[1].strip() if len(lines) > 1 else ""
# 过滤掉包含指定关键词的function_name
should_include = False
if function_name:
for exclude_name in include_function_name:
if exclude_name in function_name:
should_include = True
break
if should_include:
final_messages.append({
"role": FUNCTION,
"name": function_name,
"content": response_content
})
elif current_tag == "TOOL_CALL":
# 解析 TOOL_CALL 格式:[TOOL_CALL] function_name\narguments
lines = text.split('\n', 1)
function_name = lines[0].strip() if lines else ""
arguments = lines[1].strip() if len(lines) > 1 else ""
# 过滤掉包含指定关键词的function_name
should_include = False
if function_name:
for exclude_name in include_function_name:
if exclude_name in function_name:
should_include = True
break
if should_include:
final_messages.append({
"role": ASSISTANT,
"content": "",
"function_call": {
"name": function_name,
"arguments": arguments
}
})
elif current_tag != "THINK" and current_tag != "PREAMBLE":
final_messages.append({
"role": ASSISTANT,
"content": text
})
# if current_tag == "TOOL_RESPONSE":
# # 解析 TOOL_RESPONSE 格式:[TOOL_RESPONSE] function_name\ncontent
# lines = text.split('\n', 1)
# function_name = lines[0].strip() if lines else ""
# response_content = lines[1].strip() if len(lines) > 1 else ""
# final_messages.append({
# "role": FUNCTION,
# "name": function_name,
# "content": response_content
# })
# elif current_tag == "TOOL_CALL":
# # 解析 TOOL_CALL 格式:[TOOL_CALL] function_name\narguments
# lines = text.split('\n', 1)
# function_name = lines[0].strip() if lines else ""
# arguments = lines[1].strip() if len(lines) > 1 else ""
# final_messages.append({
# "role": ASSISTANT,
# "content": "",
# "function_call": {
# "name": function_name,
# "arguments": arguments
# }
# })
# elif current_tag != "THINK" and current_tag != "PREAMBLE":
# final_messages.append({
# "role": ASSISTANT,
# "content": text
# })
else: # 标签
current_tag = parts[i]
else:
# 非 assistant 消息或不包含 [TOOL_RESPONSE] 的消息直接添加
final_messages.append(msg)
print(final_messages)
return final_messages