feat: Add top_logprobs parameter to Responses API

client-sdks/stainless/openapi.yml

@@ -7472,6 +7472,10 @@ components:
               type: string
             type: object
           - type: 'null'
+        top_logprobs:
+          anyOf:
+          - type: integer
+          - type: 'null'
       type: object
       required:
       - input

docs/static/deprecated-llama-stack-spec.yaml

@@ -4307,6 +4307,10 @@ components:
               type: string
             type: object
           - type: 'null'
+        top_logprobs:
+          anyOf:
+          - type: integer
+          - type: 'null'
       type: object
       required:
       - input

docs/static/llama-stack-spec.yaml

@@ -6114,6 +6114,10 @@ components:
               type: string
             type: object
           - type: 'null'
+        top_logprobs:
+          anyOf:
+          - type: integer
+          - type: 'null'
       type: object
       required:
       - input

docs/static/stainless-llama-stack-spec.yaml

@@ -7472,6 +7472,10 @@ components:
               type: string
             type: object
           - type: 'null'
+        top_logprobs:
+          anyOf:
+          - type: integer
+          - type: 'null'
       type: object
       required:
       - input

src/llama_stack/providers/inline/agents/meta_reference/agents.py

@@ -113,13 +113,15 @@ async def create_openai_response(
         guardrails: list[ResponseGuardrail] | None = None,
         max_tool_calls: int | None = None,
         metadata: dict[str, str] | None = None,
+        top_logprobs: int | None = None,
     ) -> OpenAIResponseObject:
         assert self.openai_responses_impl is not None, "OpenAI responses not initialized"
         result = await self.openai_responses_impl.create_openai_response(
             input,
             model,
             prompt,
             instructions,
+            parallel_tool_calls,
             previous_response_id,
             conversation,
             store,
@@ -131,9 +133,9 @@ async def create_openai_response(
             include,
             max_infer_iters,
             guardrails,
-            parallel_tool_calls,
             max_tool_calls,
             metadata,
+            top_logprobs,
         )
         return result  # type: ignore[no-any-return]
 

src/llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py

@@ -328,6 +328,7 @@ async def create_openai_response(
         model: str,
         prompt: OpenAIResponsePrompt | None = None,
         instructions: str | None = None,
+        parallel_tool_calls: bool | None = True,
         previous_response_id: str | None = None,
         conversation: str | None = None,
         store: bool | None = True,
@@ -339,9 +340,9 @@ async def create_openai_response(
         include: list[ResponseItemInclude] | None = None,
         max_infer_iters: int | None = 10,
         guardrails: list[str | ResponseGuardrailSpec] | None = None,
-        parallel_tool_calls: bool | None = None,
         max_tool_calls: int | None = None,
         metadata: dict[str, str] | None = None,
+        top_logprobs: int | None = None,
     ):
         stream = bool(stream)
         text = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")) if text is None else text
@@ -399,6 +400,7 @@ async def create_openai_response(
             max_tool_calls=max_tool_calls,
             metadata=metadata,
             include=include,
+            top_logprobs=top_logprobs,
         )
 
         if stream:
@@ -454,6 +456,7 @@ async def _create_streaming_response(
         max_tool_calls: int | None = None,
         metadata: dict[str, str] | None = None,
         include: list[ResponseItemInclude] | None = None,
+        top_logprobs: int | None = None,
     ) -> AsyncIterator[OpenAIResponseObjectStream]:
         # These should never be None when called from create_openai_response (which sets defaults)
         # but we assert here to help mypy understand the types
@@ -505,6 +508,7 @@ async def _create_streaming_response(
             max_tool_calls=max_tool_calls,
             metadata=metadata,
             include=include,
+            top_logprobs=top_logprobs,
         )
 
         # Stream the response

src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py

@@ -138,6 +138,7 @@ def __init__(
         max_tool_calls: int | None = None,
         metadata: dict[str, str] | None = None,
         include: list[ResponseItemInclude] | None = None,
+        top_logprobs: int | None = None,
     ):
         self.inference_api = inference_api
         self.ctx = ctx
@@ -157,6 +158,7 @@ def __init__(
         self.max_tool_calls = max_tool_calls
         self.metadata = metadata
         self.include = include
+        self.top_logprobs = top_logprobs
         self.sequence_number = 0
         # Store MCP tool mapping that gets built during tool processing
         self.mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP] = (
@@ -311,6 +313,8 @@ async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
                     True if self.include and ResponseItemInclude.message_output_text_logprobs in self.include else None
                 )
 
+                top_logprobs_param = self.top_logprobs if logprobs is True else None
+
                 params = OpenAIChatCompletionRequestWithExtraBody(
                     model=self.ctx.model,
                     messages=messages,
@@ -324,6 +328,7 @@ async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
                         "include_usage": True,
                     },
                     logprobs=logprobs,
+                    top_logprobs=top_logprobs_param,
                 )
                 completion_result = await self.inference_api.openai_chat_completion(params)
 

src/llama_stack_api/agents.py

@@ -107,6 +107,7 @@ async def create_openai_response(
         ] = None,
         max_tool_calls: int | None = None,
         metadata: dict[str, str] | None = None,
+        top_logprobs: int | None = None,
     ) -> OpenAIResponseObject | AsyncIterator[OpenAIResponseObjectStream]:
         """Create a model response.
 
@@ -119,6 +120,7 @@ async def create_openai_response(
         :param guardrails: (Optional) List of guardrails to apply during response generation. Can be guardrail IDs (strings) or guardrail specifications.
         :param max_tool_calls: (Optional) Max number of total calls to built-in tools that can be processed in a response.
         :param metadata: (Optional) Dictionary of metadata key-value pairs to attach to the response.
+        :param top_logprobs: (Optional) The number of top log probabilities to return for each token. Only used when logprobs are requested via the include parameter.
         :returns: An OpenAIResponseObject.
         """
         ...