Merge pull request #4470 from pipecat-ai/pk/gpt-realtime-2-reasoning-effort

Add reasoning support to OpenAIRealtimeLLMService for gpt-realtime-2
This commit is contained in:
kompfner
2026-05-12 14:43:39 -04:00
committed by GitHub
5 changed files with 316 additions and 1 deletions

View File

@@ -164,6 +164,19 @@ class AudioConfiguration(BaseModel):
output: AudioOutput | None = None
class Reasoning(BaseModel):
"""Reasoning configuration for reasoning-capable Realtime models (e.g. ``gpt-realtime-2``).
Parameters:
effort: How much reasoning effort the model should apply. ``None``
(the default) leaves the field unset and lets the server pick.
"""
# ``| str`` for forward compatibility: if OpenAI adds new effort levels,
# users can pass the new string without waiting for a Pipecat release.
effort: Literal["minimal", "low", "medium", "high", "xhigh"] | str | None = None
class SessionProperties(BaseModel):
"""Configuration properties for an OpenAI Realtime session.
@@ -184,6 +197,8 @@ class SessionProperties(BaseModel):
prompt: Reference to a prompt template and its variables.
expires_at: Session expiration timestamp.
include: Additional fields to include in server outputs.
reasoning: Reasoning configuration. Only supported by reasoning-capable
Realtime models such as ``gpt-realtime-2``.
"""
# Needed to support ToolSchema in tools field.
@@ -206,6 +221,7 @@ class SessionProperties(BaseModel):
prompt: dict | None = None
expires_at: int | None = None
include: list[str] | None = None
reasoning: Reasoning | None = None
#

View File

@@ -321,6 +321,10 @@ class OpenAIRealtimeLLMService(LLMService[OpenAIRealtimeLLMAdapter]):
self._messages_added_manually = {}
self._pending_function_calls = {} # Track function calls by call_id
self._completed_tool_calls = set()
# Whether we've already emitted the "stripping `reasoning`" warning
# for this service instance. The Realtime API doesn't allow swapping
# the model mid-session, so once is enough.
self._reasoning_strip_warned = False
self._register_event_handler("on_conversation_item_created")
self._register_event_handler("on_conversation_item_updated")
@@ -658,6 +662,32 @@ class OpenAIRealtimeLLMService(LLMService[OpenAIRealtimeLLMAdapter]):
self._warn_unhandled_updated_settings(changed.keys() - handled)
return changed
# Substrings used to recognize reasoning-capable Realtime models. Substring
# match (rather than exact equality) so date-versioned variants of the same
# base model also match without code changes. Extend this tuple as OpenAI
# ships more reasoning-capable Realtime models.
_REASONING_CAPABLE_MODEL_SUBSTRINGS = ("gpt-realtime-2",)
def _strip_unsupported_reasoning(
self, settings: events.SessionProperties
) -> events.SessionProperties:
"""Drop ``reasoning`` from an outgoing session.update if the model can't use it.
The server otherwise rejects the whole update and kills the session.
Returns a copy when stripping; the user's stored config is preserved.
"""
if settings.reasoning is None or not settings.model:
return settings
if any(s in settings.model for s in self._REASONING_CAPABLE_MODEL_SUBSTRINGS):
return settings
if not self._reasoning_strip_warned:
logger.warning(
f"{self} stripping `reasoning` from session.update: model={settings.model!r} "
f"isn't a known reasoning-capable Realtime model."
)
self._reasoning_strip_warned = True
return settings.model_copy(update={"reasoning": None})
async def _send_session_update(self):
settings = assert_given(self._settings.session_properties)
adapter = self.get_llm_adapter()
@@ -683,7 +713,9 @@ class OpenAIRealtimeLLMService(LLMService[OpenAIRealtimeLLMAdapter]):
if settings.tools and isinstance(settings.tools, ToolsSchema):
settings.tools = adapter.from_standard_tools(settings.tools)
await self.send_client_event(events.SessionUpdateEvent(session=settings))
outgoing = self._strip_unsupported_reasoning(settings)
await self.send_client_event(events.SessionUpdateEvent(session=outgoing))
#
# inbound server event handling