phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-13 13:22:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add support for instructions parameter in response object

Browse source
This commit is contained in:
Shabana Baig 2025-10-14 14:39:23 -04:00
parent 08cbb69ef7
commit f176e1a74b
10 changed files with 229 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

12
llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py
View file

@ -123,6 +123,17 @@ class OpenAIResponsesImpl:
# Use stored messages directly and convert only new input
message_adapter = TypeAdapter(list[OpenAIMessageParam])
messages = message_adapter.validate_python(previous_response.messages)
# When managing conversation state with the previous_response_id parameter,
# the instructions used on previous turns will not be carried over in the context
previous_instructions = previous_response.instructions
if previous_instructions:
if (isinstance(previous_instructions, str) and
previous_instructions == messages[0].content and
messages[0].role == "system"):
# Omit instructions from previous response
del messages[0]
else:
raise ValueError("Instructions from the previous response could not be validated")
new_messages = await convert_response_input_to_chat_messages(input, previous_messages=messages)
messages.extend(new_messages)
else:
@ -359,6 +370,7 @@ class OpenAIResponsesImpl:
tool_executor=self.tool_executor,
safety_api=self.safety_api,
guardrail_ids=guardrail_ids,
instructions=instructions,
)
# Stream the response

4
llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
View file

@ -112,6 +112,7 @@ class StreamingResponseOrchestrator:
tool_executor, # Will be the tool execution logic from the main class
safety_api,
guardrail_ids: list[str] | None = None,
instructions: str,
):
self.inference_api = inference_api
self.ctx = ctx
@ -133,6 +134,8 @@ class StreamingResponseOrchestrator:
self.accumulated_usage: OpenAIResponseUsage | None = None
# Track if we've sent a refusal response
self.violation_detected = False
# system message that is inserted into the model's context
self.instructions = instructions
async def _create_refusal_response(self, violation_message: str) -> OpenAIResponseObjectStream:
"""Create a refusal response to replace streaming content."""
@ -176,6 +179,7 @@ class StreamingResponseOrchestrator:
tools=self.ctx.available_tools(),
error=error,
usage=self.accumulated_usage,
instructions=self.instructions,
)
async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]: