fix: empty or unreadable replies for API clients on a Responses API connection

Apps calling Open WebUI with an API key, for example through the OpenAI SDK, could not use replies from a model whose connection is set to the Responses API. The SDK rejected normal replies because the token counts and creation time were missing, and streamed replies came through empty, also on the Anthropic-style /api/v1/messages endpoint. A reply that was not streamed and got cut off at the length limit or by a content filter was reported as finished normally, so the app could not tell it was cut off. Replies now carry the token counts, creation time and finish reason the SDK expects, and streamed replies arrive in the same format as from any other connection. Streams are converted where the reply goes out to the API client, because for some models, such as arena models, the real connection is only known once the reply is streaming.

Fixes #33389
Fixes #33445
This commit is contained in:
Classic298 2026-10-30 09:45:05 +01:00
parent 63312766f8
commit c52e78cc7f
3 changed files with 113 additions and 3 deletions

View file

@ -4,6 +4,7 @@ import asyncio
import hashlib
import logging
import re
import time
from typing import Optional
from urllib.parse import quote, urlparse
@ -48,6 +49,7 @@ from open_webui.utils.payload import (
apply_model_params_to_body_openai,
apply_system_prompt_to_body,
)
from open_webui.utils.response import convert_responses_usage_to_openai
from open_webui.utils.session_pool import (
cleanup_response,
get_client_timeout,
@ -1384,9 +1386,16 @@ def convert_responses_result(response: dict) -> dict:
}
)
if response.get('status') == 'incomplete':
incomplete_reason = (response.get('incomplete_details') or {}).get('reason')
finish_reason = 'content_filter' if incomplete_reason == 'content_filter' else 'length'
else:
finish_reason = 'tool_calls' if tool_calls else 'stop'
return {
'id': response.get('id', ''),
'object': 'chat.completion',
'created': int(time.time()),
'model': response.get('model', ''),
'choices': [
{
@ -1396,10 +1405,10 @@ def convert_responses_result(response: dict) -> dict:
'content': content,
**({'tool_calls': tool_calls} if tool_calls else {}),
},
'finish_reason': 'tool_calls' if tool_calls else 'stop',
'finish_reason': finish_reason,
}
],
'usage': response.get('usage', {}),
'usage': convert_responses_usage_to_openai(response.get('usage')),
}

View file

@ -131,6 +131,7 @@ from open_webui.utils.oauth import get_system_oauth_token
from open_webui.utils.payload import apply_params_to_form_data, apply_system_prompt_to_body, resolve_system_prompt
from open_webui.utils.plugin import load_function_module_by_id
from open_webui.utils.response import (
convert_streaming_response_responses_to_openai,
get_response_delta_key,
handle_responses_streaming_event,
merge_usage,
@ -6534,7 +6535,7 @@ async def streaming_chat_response_handler(response, ctx):
log_model_usage(request, user, form_data['model'], current_usage)
return StreamingResponse(
stream_wrapper(response.body_iterator, events),
stream_wrapper(convert_streaming_response_responses_to_openai(response.body_iterator), events),
headers=dict(response.headers),
background=response.background,
)

View file

@ -1,4 +1,5 @@
from numbers import Number
from typing import AsyncIterator
from uuid import uuid4
from open_webui.utils.json_codec import JSONCodec
@ -231,6 +232,16 @@ def convert_ollama_usage_to_openai(data: dict) -> dict:
return result
def convert_responses_usage_to_openai(usage: dict | None) -> dict:
if not usage:
return {}
return {
**usage,
'prompt_tokens': usage.get('input_tokens', 0),
'completion_tokens': usage.get('output_tokens', 0),
}
def convert_response_ollama_to_openai(ollama_response: dict) -> dict:
model = ollama_response.get('model', 'ollama')
message_content = ollama_response.get('message', {}).get('content', '')
@ -299,6 +310,95 @@ async def convert_streaming_response_ollama_to_openai(ollama_streaming_response)
yield 'data: [DONE]\n\n'
async def convert_streaming_response_responses_to_openai(body_iterator: AsyncIterator) -> AsyncIterator[str | bytes]:
"""Translate Responses API events into Chat Completions chunks; other lines pass through."""
completion_id = f'chatcmpl-{str(uuid4())}'
model = ''
tool_call_indexes = {}
first = True
is_responses_stream = False
async for line in body_iterator:
text = line.decode('utf-8', 'replace') if isinstance(line, bytes) else line
# Event names would mislabel the translated chunks.
if text.startswith('event:') and text.removeprefix('event:').strip().startswith('response.'):
continue
try:
data = JSONCodec.loads(text.removeprefix('data:').strip()) if text.startswith('data:') else None
except JSONCodec.JSONDecodeError:
data = None
event_type = data.get('type', '') if isinstance(data, dict) else ''
if not event_type.startswith('response.') and not (event_type == 'error' and is_responses_stream):
yield line
continue
if event_type in ('response.failed', 'error'):
error = data.get('response', {}).get('error', {}) if event_type == 'response.failed' else data
yield f'data: {JSONCodec.dumps({"error": error})}\n\n'
yield 'data: [DONE]\n\n'
continue
if event_type == 'response.created':
model = data.get('response', {}).get('model', '')
is_responses_stream = True
continue
output_index = data.get('output_index')
delta = data.get('delta')
is_final = event_type in ('response.completed', 'response.incomplete')
content = None
reasoning_content = None
tool_calls = None
usage = None
if event_type == 'response.output_text.delta' and delta:
content = delta
elif event_type in ('response.reasoning_text.delta', 'response.reasoning_summary_text.delta') and delta:
reasoning_content = delta
elif event_type == 'response.output_item.added' and data.get('item', {}).get('type') == 'function_call':
item = data['item']
tool_call_indexes[output_index] = len(tool_call_indexes)
tool_calls = [
{
'index': tool_call_indexes[output_index],
'id': item.get('call_id', ''),
'type': 'function',
'function': {'name': item.get('name', ''), 'arguments': item.get('arguments', '')},
}
]
elif event_type == 'response.function_call_arguments.delta':
tool_calls = [
{
'index': tool_call_indexes.get(output_index, 0),
'function': {'arguments': delta},
}
]
elif is_final:
usage = convert_responses_usage_to_openai(data.get('response', {}).get('usage'))
else:
continue
chunk = openai_chat_chunk_message_template(
model, content, reasoning_content, tool_calls, usage, message_id=completion_id
)
if first:
chunk['choices'][0]['delta']['role'] = 'assistant'
first = False
if event_type == 'response.incomplete':
incomplete_reason = (data.get('response', {}).get('incomplete_details') or {}).get('reason')
chunk['choices'][0]['finish_reason'] = (
'content_filter' if incomplete_reason == 'content_filter' else 'length'
)
elif is_final and tool_call_indexes:
chunk['choices'][0]['finish_reason'] = 'tool_calls'
yield f'data: {JSONCodec.dumps(chunk)}\n\n'
if is_final:
yield 'data: [DONE]\n\n'
def convert_embedding_response_ollama_to_openai(response) -> dict:
"""
Convert the response from Ollama embeddings endpoint to the OpenAI-compatible format.