OpenAI Responses API Plugin #4192
Conversation
| if isinstance(event, ResponseErrorEvent): | ||
| self._handle_error(event) | ||
|
|
||
| if isinstance(event, ResponseCreatedEvent): | ||
| self._handle_response_created(event) | ||
|
|
||
| if isinstance(event, ResponseOutputItemDoneEvent): | ||
| chunk = self._handle_output_items_done(event) | ||
|
|
||
| if isinstance(event, ResponseTextDeltaEvent): | ||
| chunk = self._handle_response_output_text_delta(event) | ||
|
|
||
| if isinstance(event, ResponseCompletedEvent): | ||
| chunk = self._handle_response_completed(event) | ||
|
|
||
| if chunk is not None: | ||
| self._event_ch.send_nowait(chunk) |
There was a problem hiding this comment.
Should we set verbosity to 'low' or None and disable thinking by default? This mainly applies when working with the GPT-5 family models. Otherwise, the time to first text token tends to be quite slow, which can make it feel like the LLM request is hanging.
There was a problem hiding this comment.
The alternative would be streaming reasoning token as llm chunks (user can set weather to let LLM "speak out" reasoning token with a flag), or we do stream thinking chunk inside LLM plugin but make it only sends as part of transcripts but not pass into TTS. cc @theomonnom let me know your thoughts
| elif isinstance(tool, llm.ProviderTool): | ||
| schemas.append(tool.to_dict()) |
There was a problem hiding this comment.
We should manually check every ProviderTool individually here (FileSearch, ...).
llm.ProviderTool doesn't have a to_dict() method.
There was a problem hiding this comment.
just updated, checking openai.tools.OpenAITool works
|
woohoo! nice work @tinalenguyen |
To use:
llm=openai.responses.LLM()The tools included:
openai.tools.WebSearch()openai.tools.FileSearch()openai.tools.CodeInterpreter()