xinyin025/LangBot
1
0
Fork 0
mirror of https://github.com/langbot-app/LangBot.git synced 2026-06-17 11:14:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix:The handling of the streaming tool calls has been fixed, but there are still bugs in the model's reply messages with thoughtfulness.

Browse Source
This commit is contained in:
Dong_master
2025-08-10 23:14:57 +08:00
parent 03b11481ed
commit 46452de7b5
10 changed files with 112 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pkg/provider/modelmgr/requesters/anthropicmsgs.py
+29 -8
View File
@@ -273,13 +273,11 @@ class AnthropicMessages(requester.ProviderAPIRequester):
del msg_dict['tool_calls']
req_messages.append(msg_dict)
if args["thinking"]:
if args.get("thinking", False):
args['thinking'] = {
"type": "enabled",
"budget_tokens": 10000
}
else:
args.pop('thinking')
args['messages'] = req_messages
@@ -296,15 +294,32 @@ class AnthropicMessages(requester.ProviderAPIRequester):
think_ended = False
finish_reason = False
content = ''
tool_name = ''
tool_id = ''
tool_calls = []
async for chunk in await self.client.messages.create(**args):
# print(chunk)
print(chunk)
tool_call = {"id":None, 'function': {"name": None, "arguments": None},'type':'function'}
if isinstance(chunk, anthropic.types.raw_content_block_start_event.RawContentBlockStartEvent): # 记录开始
if chunk.content_block.type == 'tool_use':
if chunk.content_block.name is not None:
tool_name = chunk.content_block.name
if chunk.content_block.id is not None:
tool_id = chunk.content_block.id
tool_call['function']['name'] = tool_name
tool_call['function']['arguments'] = ''
tool_call['id'] = tool_id
print(chunk.content_block)
if not remove_think:
if chunk.content_block.type == 'thinking':
if chunk.content_block.type == 'thinking' and not remove_think:
think_started = True
elif chunk.content_block.type == 'text':
elif chunk.content_block.type == 'text' and chunk.index != 0 and not remove_think:
think_ended = True
continue
continue
elif isinstance(chunk, anthropic.types.raw_content_block_delta_event.RawContentBlockDeltaEvent):
if chunk.delta.type == "thinking_delta":
if think_started:
@@ -320,6 +335,10 @@ class AnthropicMessages(requester.ProviderAPIRequester):
content = '\n</think>\n' + chunk.delta.text
else:
content = chunk.delta.text
elif chunk.delta.type == "input_json_delta":
tool_call['function']["arguments"] = chunk.delta.partial_json
tool_call['function']['name'] = tool_name
tool_call['id'] = tool_id
elif isinstance(chunk, anthropic.types.raw_content_block_stop_event.RawContentBlockStopEvent):
continue # 记录raw_content_block结束的
@@ -333,10 +352,12 @@ class AnthropicMessages(requester.ProviderAPIRequester):
continue
args = {
'content': content,
'role': role,
"is_final": finish_reason
"is_final": finish_reason,
'tool_calls': None if tool_call['id'] is None else [tool_call],
}
# if chunk_idx == 0:
# chunk_idx += 1
pkg/provider/modelmgr/requesters/chatcmpl.py
+16 -26
View File
@@ -160,18 +160,21 @@ class OpenAIChatCompletions(requester.ProviderAPIRequester):
thinking_started = False
thinking_ended = False
role = 'assistant' # 默认角色
tool_id = ""
tool_name = ''
# accumulated_reasoning = '' # 仅用于判断何时结束思维链
async for chunk in self._req_stream(args, extra_body=extra_args):
# 解析 chunk 数据
if hasattr(chunk, 'choices') and chunk.choices:
choice = chunk.choices[0]
delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
finish_reason = getattr(choice, 'finish_reason', None)
else:
delta = {}
finish_reason = None
# 从第一个 chunk 获取 role，后续使用这个 role
if 'role' in delta and delta['role']:
role = delta['role']
@@ -208,41 +211,29 @@ class OpenAIChatCompletions(requester.ProviderAPIRequester):
# delta_content = re.sub(r'<think>.*?</think>', '', delta_content, flags=re.DOTALL)
# 处理工具调用增量
delta_tool_calls = None
# delta_tool_calls = None
if delta.get('tool_calls'):
delta_tool_calls = []
for tool_call in delta['tool_calls']:
tc_id = tool_call.get('id')
if tc_id:
if tc_id not in tool_calls_map:
# 新的工具调用
tool_calls_map[tc_id] = llm_entities.ToolCall(
id=tc_id,
type=tool_call.get('type', 'function'),
function=llm_entities.FunctionCall(
name=tool_call.get('function', {}).get('name', ''),
arguments=tool_call.get('function', {}).get('arguments', ''),
),
)
delta_tool_calls.append(tool_calls_map[tc_id])
else:
# 追加函数参数
func_args = tool_call.get('function', {}).get('arguments', '')
if func_args:
tool_calls_map[tc_id].function.arguments += func_args
# 返回更新后的完整工具调用
delta_tool_calls.append(tool_calls_map[tc_id])
if tool_call['id'] and tool_call['function']['name']:
tool_id = tool_call['id']
tool_name = tool_call['function']['name']
else:
tool_call['id'] = tool_id
tool_call['function']['name'] = tool_name
if tool_call['type'] is None:
tool_call['type'] = 'function'
# 跳过空的第一个 chunk（只有 role 没有内容）
if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
chunk_idx += 1
continue
# 构建 MessageChunk - 只包含增量内容
chunk_data = {
'role': role,
'content': delta_content if delta_content else None,
'tool_calls': delta_tool_calls if delta_tool_calls else None,
'tool_calls': delta.get('tool_calls'),
'is_final': bool(finish_reason),
}
@@ -289,7 +280,6 @@ class OpenAIChatCompletions(requester.ProviderAPIRequester):
# 发送请求
resp = await self._req(args, extra_body=extra_args)
print(resp)
# 处理请求结果
message = await self._make_msg(resp, remove_think)
pkg/provider/modelmgr/requesters/modelscopechatcmpl.py
+9 -23
View File
@@ -289,30 +289,16 @@ class ModelScopeChatCompletions(requester.ProviderAPIRequester):
# delta_content = re.sub(r'<think>.*?</think>', '', delta_content, flags=re.DOTALL)
# 处理工具调用增量
delta_tool_calls = None
if delta.get('tool_calls'):
delta_tool_calls = []
for tool_call in delta['tool_calls']:
tc_id = tool_call.get('id')
if tc_id:
if tc_id not in tool_calls_map:
# 新的工具调用
tool_calls_map[tc_id] = llm_entities.ToolCall(
id=tc_id,
type=tool_call.get('type', 'function'),
function=llm_entities.FunctionCall(
name=tool_call.get('function', {}).get('name', ''),
arguments=tool_call.get('function', {}).get('arguments', ''),
),
)
delta_tool_calls.append(tool_calls_map[tc_id])
else:
# 追加函数参数
func_args = tool_call.get('function', {}).get('arguments', '')
if func_args:
tool_calls_map[tc_id].function.arguments += func_args
# 返回更新后的完整工具调用
delta_tool_calls.append(tool_calls_map[tc_id])
if tool_call['id'] and tool_call['function']['name']:
tool_id = tool_call['id']
tool_name = tool_call['function']['name']
else:
tool_call['id'] = tool_id
tool_call['function']['name'] = tool_name
if tool_call['type'] is None:
tool_call['type'] = 'function'
# 跳过空的第一个 chunk（只有 role 没有内容）
if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
@@ -323,7 +309,7 @@ class ModelScopeChatCompletions(requester.ProviderAPIRequester):
chunk_data = {
'role': role,
'content': delta_content if delta_content else None,
'tool_calls': delta_tool_calls if delta_tool_calls else None,
'tool_calls': delta.get('tool_calls'),
'is_final': bool(finish_reason),
}
pkg/provider/modelmgr/requesters/ppiochatcmpl.py
+14 -23
View File
@@ -179,39 +179,30 @@ class PPIOChatCompletions(chatcmpl.OpenAIChatCompletions):
delta_tool_calls = None
if delta.get('tool_calls'):
delta_tool_calls = []
for tool_call in delta['tool_calls']:
tc_id = tool_call.get('id')
if tc_id:
if tc_id not in tool_calls_map:
# 新的工具调用
tool_calls_map[tc_id] = llm_entities.ToolCall(
id=tc_id,
type=tool_call.get('type', 'function'),
function=llm_entities.FunctionCall(
name=tool_call.get('function', {}).get('name', ''),
arguments=tool_call.get('function', {}).get('arguments', ''),
),
)
delta_tool_calls.append(tool_calls_map[tc_id])
else:
# 追加函数参数
func_args = tool_call.get('function', {}).get('arguments', '')
if func_args:
tool_calls_map[tc_id].function.arguments += func_args
# 返回更新后的完整工具调用
delta_tool_calls.append(tool_calls_map[tc_id])
if tool_call['id'] and tool_call['function']['name']:
tool_id = tool_call['id']
tool_name = tool_call['function']['name']
if tool_call['id'] is None:
tool_call['id'] = tool_id
if tool_call['function']['name'] is None:
tool_call['function']['name'] = tool_name
if tool_call['function']['arguments'] is None:
tool_call['function']['arguments'] = ''
if tool_call['type'] is None:
tool_call['type'] = 'function'
# 跳过空的第一个 chunk（只有 role 没有内容）
if chunk_idx == 0 and not delta_content and not delta.get('tool_calls'):
chunk_idx += 1
continue
# 构建 MessageChunk - 只包含增量内容
# 构建 MessageChunk - 只包含增量内容
chunk_data = {
'role': role,
'content': delta_content if delta_content else None,
'tool_calls': delta_tool_calls if delta_tool_calls else None,
'tool_calls': delta.get('tool_calls'),
'is_final': bool(finish_reason),
}
pkg/provider/runners/localagent.py
+14 -8
View File
@@ -148,13 +148,13 @@ class LocalAgentRunner(runner.RequestRunner):
if tool_call.function and tool_call.function.arguments:
# 流式处理中，工具调用参数可能分多个chunk返回，需要追加而不是覆盖
tool_calls_map[tool_call.id].function.arguments += tool_call.function.arguments
print(list(tool_calls_map.values()) if (tool_calls_map and msg.is_final) else None)
# continue
# 每8个chunk或最后一个chunk时，输出所有累积的内容
if msg_idx % 8 == 0 or msg.is_final:
yield llm_entities.MessageChunk(
role=last_role,
content=accumulated_content, # 输出所有累积内容
tool_calls=list(tool_calls_map.values()) if (tool_calls_map and msg.is_final) else None,
is_final=msg.is_final,
)
@@ -178,12 +178,18 @@ class LocalAgentRunner(runner.RequestRunner):
parameters = json.loads(func.arguments)
func_ret = await self.ap.tool_mgr.execute_func_call(query, func.name, parameters)
msg = llm_entities.Message(
role='tool',
content=json.dumps(func_ret, ensure_ascii=False),
tool_call_id=tool_call.id,
)
if is_stream:
msg = llm_entities.MessageChunk(
role='tool',
content=json.dumps(func_ret, ensure_ascii=False),
tool_call_id=tool_call.id,
)
else:
msg = llm_entities.Message(
role='tool',
content=json.dumps(func_ret, ensure_ascii=False),
tool_call_id=tool_call.id,
)
yield msg