mirror of
https://github.com/open-webui/open-webui.git
synced 2026-09-16 23:43:03 +00:00
fix(mcp): execute MCP tool calls server-side in native function calling mode
When Ollama/Gemma returns finish_reason=tool_calls in a non-streaming response, OWU was passing the response through unchanged — the tool was never called and oracle-mcp only ever saw ListToolsRequest. Root cause: the tool-execution loop lives entirely in streaming_chat_response_handler, which is skipped for non-streaming JSON responses. non_streaming_chat_response_handler only checked for choices[0].message.content and silently dropped tool_calls. Fix: - Add _execute_tool_calls_non_streaming() helper that extracts tool calls from a non-streaming LLM response, resolves each tool from metadata['tools'], executes it (MCP call_tool or builtin callable), and feeds the results back for a second LLM turn. - Call it from non_streaming_chat_response_handler whenever function_calling==native and tool_calls are present in the response. - Also normalize Ollama-prefixed tool names (e.g. "oracle:oracle_vault_ask" → "oracle_vault_ask") at execution time in both the streaming and non-streaming paths so the tools_dict lookup succeeds. Tested: oracle-mcp now shows CallToolRequest after Gemma emits a tool_call, and Gemma's final answer is grounded in vault data. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Prometheus
parent
8dae237a0b
commit
d6eb4544fb
1 changed files with 173 additions and 0 deletions
|
|
@ -3342,6 +3342,144 @@ async def outlet_filter_handler(ctx):
|
|||
log.debug(f'Error running outlet filters: {e}')
|
||||
|
||||
|
||||
async def _execute_tool_calls_non_streaming(request, form_data, ctx, response_data):
|
||||
"""Execute tool calls from a non-streaming LLM response and return the final response.
|
||||
|
||||
When a model returns finish_reason=tool_calls in a non-streaming response,
|
||||
OWU needs to:
|
||||
1. Execute each requested tool (MCP or builtin)
|
||||
2. Append the assistant message and tool results to the message history
|
||||
3. Issue a second LLM call and return its response
|
||||
"""
|
||||
user = ctx['user']
|
||||
metadata = ctx['metadata']
|
||||
|
||||
tools = metadata.get('tools', {})
|
||||
if not tools:
|
||||
return None
|
||||
|
||||
choices = response_data.get('choices', [])
|
||||
if not choices:
|
||||
return None
|
||||
|
||||
message = choices[0].get('message', {})
|
||||
tool_calls_raw = message.get('tool_calls', [])
|
||||
if not tool_calls_raw:
|
||||
return None
|
||||
|
||||
results = []
|
||||
for tool_call in tool_calls_raw:
|
||||
tool_call_id = tool_call.get('id', '')
|
||||
tool_function_name = tool_call.get('function', {}).get('name', '')
|
||||
tool_args = tool_call.get('function', {}).get('arguments', '{}')
|
||||
|
||||
# Normalize tool name: Ollama may prepend the MCP server id as a
|
||||
# namespace (e.g. "oracle:oracle_vault_ask") but tools_dict keys use
|
||||
# the underscore-prefixed form ("oracle_vault_ask").
|
||||
if tool_function_name not in tools and ':' in tool_function_name:
|
||||
normalized_name = tool_function_name.split(':', 1)[1]
|
||||
if normalized_name in tools:
|
||||
log.info(f'[MCP] Normalizing tool name "{tool_function_name}" -> "{normalized_name}"')
|
||||
tool_function_name = normalized_name
|
||||
tool_call.setdefault('function', {})['name'] = tool_function_name
|
||||
|
||||
# Parse arguments
|
||||
tool_function_params = {}
|
||||
if tool_args and tool_args.strip():
|
||||
try:
|
||||
tool_function_params = ast.literal_eval(tool_args)
|
||||
except Exception:
|
||||
try:
|
||||
tool_function_params = json.loads(tool_args)
|
||||
except Exception:
|
||||
log.error(f'[MCP] Could not parse tool args for {tool_function_name}: {tool_args}')
|
||||
results.append({'tool_call_id': tool_call_id, 'content': 'Error: could not parse arguments'})
|
||||
continue
|
||||
|
||||
# Sanitize args to only allowed params
|
||||
tool_result = None
|
||||
if tool_function_name in tools:
|
||||
tool = tools[tool_function_name]
|
||||
spec = tool.get('spec', {})
|
||||
allowed_params = spec.get('parameters', {}).get('properties', {}).keys()
|
||||
tool_function_params = {k: v for k, v in tool_function_params.items() if k in allowed_params}
|
||||
|
||||
try:
|
||||
log.info(f'[MCP] Executing tool "{tool_function_name}" params={tool_function_params}')
|
||||
if tool.get('direct', False):
|
||||
event_caller = ctx.get('event_caller')
|
||||
if event_caller:
|
||||
tool_result = await event_caller(
|
||||
{
|
||||
'type': 'execute:tool',
|
||||
'data': {
|
||||
'id': str(uuid4()),
|
||||
'name': tool_function_name,
|
||||
'params': tool_function_params,
|
||||
'server': tool.get('server', {}),
|
||||
'session_id': metadata.get('session_id', None),
|
||||
},
|
||||
}
|
||||
)
|
||||
else:
|
||||
tool_function = await get_updated_tool_function(
|
||||
function=tool['callable'],
|
||||
extra_params={
|
||||
'__messages__': form_data.get('messages', []),
|
||||
'__files__': metadata.get('files', []),
|
||||
},
|
||||
)
|
||||
tool_result = await tool_function(**tool_function_params)
|
||||
log.info(f'[MCP] Tool "{tool_function_name}" result: {str(tool_result)[:200]}')
|
||||
except Exception as e:
|
||||
log.error(f'[MCP] Tool execution error for {tool_function_name}: {e}')
|
||||
tool_result = str(e)
|
||||
else:
|
||||
log.warning(f'[MCP] Tool "{tool_function_name}" not found in tools_dict (keys: {list(tools.keys())[:5]}...)')
|
||||
tool_result = f'Error: tool "{tool_function_name}" not found'
|
||||
|
||||
results.append({'tool_call_id': tool_call_id, 'content': str(tool_result) if tool_result else ''})
|
||||
|
||||
if not results:
|
||||
return None
|
||||
|
||||
# Build updated messages: original + assistant msg with tool_calls + tool results
|
||||
updated_messages = list(form_data.get('messages', []))
|
||||
updated_messages.append({
|
||||
'role': 'assistant',
|
||||
'content': message.get('content') or '',
|
||||
'tool_calls': tool_calls_raw,
|
||||
})
|
||||
for result in results:
|
||||
updated_messages.append({
|
||||
'role': 'tool',
|
||||
'tool_call_id': result['tool_call_id'],
|
||||
'content': result['content'],
|
||||
})
|
||||
|
||||
# Second LLM call with tool results in context
|
||||
new_form_data = {
|
||||
**form_data,
|
||||
'messages': updated_messages,
|
||||
'stream': False,
|
||||
}
|
||||
# Remove tools list so the LLM just answers (avoids infinite loop)
|
||||
new_form_data.pop('tools', None)
|
||||
|
||||
log.info(f'[MCP] Making second LLM call after tool execution')
|
||||
try:
|
||||
second_response = await generate_chat_completion(
|
||||
request,
|
||||
new_form_data,
|
||||
user,
|
||||
bypass_system_prompt=True,
|
||||
)
|
||||
return second_response
|
||||
except Exception as e:
|
||||
log.error(f'[MCP] Error in second LLM call: {e}')
|
||||
return None
|
||||
|
||||
|
||||
async def non_streaming_chat_response_handler(response, ctx):
|
||||
request = ctx['request']
|
||||
|
||||
|
|
@ -3351,10 +3489,28 @@ async def non_streaming_chat_response_handler(response, ctx):
|
|||
|
||||
event_emitter = ctx['event_emitter']
|
||||
|
||||
form_data = ctx.get('form_data', {})
|
||||
|
||||
response, response_data = get_response_data(response)
|
||||
if response_data is None:
|
||||
return response
|
||||
|
||||
# Handle tool_calls from native function calling in non-streaming responses.
|
||||
# The streaming_chat_response_handler's tool-execution loop only runs when
|
||||
# the response IS a StreamingResponse. For non-streaming JSON responses
|
||||
# (finish_reason="tool_calls"), we must execute tools here and make a second
|
||||
# LLM call with the results before returning to the caller.
|
||||
if metadata.get('params', {}).get('function_calling') == 'native':
|
||||
choices = response_data.get('choices', [])
|
||||
if choices and choices[0].get('message', {}).get('tool_calls'):
|
||||
log.info('[MCP] non-streaming tool_calls detected — executing and re-calling LLM')
|
||||
second_response = await _execute_tool_calls_non_streaming(
|
||||
request, form_data, ctx, response_data
|
||||
)
|
||||
if second_response is not None:
|
||||
# Recurse once so the second response goes through normal content handling
|
||||
return await non_streaming_chat_response_handler(second_response, ctx)
|
||||
|
||||
if event_emitter:
|
||||
try:
|
||||
if 'error' in response_data:
|
||||
|
|
@ -4461,6 +4617,20 @@ async def streaming_chat_response_handler(response, ctx):
|
|||
tool_function_name = tool_call.get('function', {}).get('name', '')
|
||||
tool_args = tool_call.get('function', {}).get('arguments', '{}')
|
||||
|
||||
# Normalize tool name: some backends (e.g. Ollama) prepend the MCP
|
||||
# server id as a namespace (e.g. "oracle:oracle_vault_ask") but our
|
||||
# tools_dict keys use only the underscore-prefixed form
|
||||
# ("oracle_vault_ask"). Strip the "server_id:" prefix so the lookup
|
||||
# below succeeds.
|
||||
if tool_function_name not in tools and ':' in tool_function_name:
|
||||
normalized_name = tool_function_name.split(':', 1)[1]
|
||||
if normalized_name in tools:
|
||||
log.info(
|
||||
f'Normalizing tool name from "{tool_function_name}" to "{normalized_name}"'
|
||||
)
|
||||
tool_function_name = normalized_name
|
||||
tool_call.setdefault('function', {})['name'] = tool_function_name
|
||||
|
||||
tool_function_params = {}
|
||||
if tool_args and tool_args.strip():
|
||||
try:
|
||||
|
|
@ -4490,6 +4660,9 @@ async def streaming_chat_response_handler(response, ctx):
|
|||
tool_type = None
|
||||
direct_tool = False
|
||||
|
||||
log.info(
|
||||
f'[MCP] tool_call: name="{tool_function_name}" found={tool_function_name in tools}'
|
||||
)
|
||||
if tool_function_name in tools:
|
||||
tool = tools[tool_function_name]
|
||||
spec = tool.get('spec', {})
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue