llama-stack-mirror/llama_stack/apis/agents/openai_responses.py
Ben Browning 8e316c9b1e
feat: function tools in OpenAI Responses (#2094)
# What does this PR do?

This is a combination of what was previously 3 separate PRs - #2069,
#2075, and #2083. It turns out all 3 of those are needed to land a
working function calling Responses implementation. The web search
builtin tool was already working, but this wires in support for custom
function calling.

I ended up combining all three into one PR because they all had lots of
merge conflicts, both with each other but also with #1806 that just
landed. And, because landing any of them individually would have only
left a partially working implementation merged.

The new things added here are:
* Storing of input items from previous responses and restoring of those
input items when adding previous responses to the conversation state
* Handling of multiple input item messages roles, not just "user"
messages.
* Support for custom tools passed into the Responses API to enable
function calling outside of just the builtin websearch tool.

Closes #2074
Closes #2080

## Test Plan

### Unit Tests

Several new unit tests were added, and they all pass. Ran via:

```
python -m pytest -s -v tests/unit/providers/agents/meta_reference/test_openai_responses.py
```

### Responses API Verification Tests

I ran our verification run.yaml against multiple providers to ensure we
were getting a decent pass rate. Specifically, I ensured the new custom
tool verification test passed across multiple providers and that the
multi-turn examples passed across at least some of the providers (some
providers struggle with the multi-turn workflows still).

Running the stack setup for verification testing:

```
llama stack run --image-type venv tests/verifications/openai-api-verification-run.yaml
```

Together, passing 100% as an example:

```
pytest -s -v 'tests/verifications/openai_api/test_responses.py' --provider=together-llama-stack
```

## Documentation

We will need to start documenting the OpenAI APIs, but for now the
Responses stuff is still rapidly evolving so delaying that.

---------

Signed-off-by: Derek Higgins <derekh@redhat.com>
Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Derek Higgins <derekh@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
2025-05-13 11:29:15 -07:00

198 lines
5.9 KiB
Python

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Annotated, Any, Literal
from pydantic import BaseModel, Field
from llama_stack.schema_utils import json_schema_type, register_schema
@json_schema_type
class OpenAIResponseError(BaseModel):
code: str
message: str
@json_schema_type
class OpenAIResponseInputMessageContentText(BaseModel):
text: str
type: Literal["input_text"] = "input_text"
@json_schema_type
class OpenAIResponseInputMessageContentImage(BaseModel):
detail: Literal["low"] | Literal["high"] | Literal["auto"] = "auto"
type: Literal["input_image"] = "input_image"
# TODO: handle file_id
image_url: str | None = None
# TODO: handle file content types
OpenAIResponseInputMessageContent = Annotated[
OpenAIResponseInputMessageContentText | OpenAIResponseInputMessageContentImage,
Field(discriminator="type"),
]
register_schema(OpenAIResponseInputMessageContent, name="OpenAIResponseInputMessageContent")
@json_schema_type
class OpenAIResponseOutputMessageContentOutputText(BaseModel):
text: str
type: Literal["output_text"] = "output_text"
OpenAIResponseOutputMessageContent = Annotated[
OpenAIResponseOutputMessageContentOutputText,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutputMessageContent, name="OpenAIResponseOutputMessageContent")
@json_schema_type
class OpenAIResponseMessage(BaseModel):
"""
Corresponds to the various Message types in the Responses API.
They are all under one type because the Responses API gives them all
the same "type" value, and there is no way to tell them apart in certain
scenarios.
"""
content: str | list[OpenAIResponseInputMessageContent] | list[OpenAIResponseOutputMessageContent]
role: Literal["system"] | Literal["developer"] | Literal["user"] | Literal["assistant"]
type: Literal["message"] = "message"
# The fields below are not used in all scenarios, but are required in others.
id: str | None = None
status: str | None = None
@json_schema_type
class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
id: str
status: str
type: Literal["web_search_call"] = "web_search_call"
@json_schema_type
class OpenAIResponseOutputMessageFunctionToolCall(BaseModel):
arguments: str
call_id: str
name: str
type: Literal["function_call"] = "function_call"
id: str
status: str
OpenAIResponseOutput = Annotated[
OpenAIResponseMessage | OpenAIResponseOutputMessageWebSearchToolCall | OpenAIResponseOutputMessageFunctionToolCall,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
@json_schema_type
class OpenAIResponseObject(BaseModel):
created_at: int
error: OpenAIResponseError | None = None
id: str
model: str
object: Literal["response"] = "response"
output: list[OpenAIResponseOutput]
parallel_tool_calls: bool = False
previous_response_id: str | None = None
status: str
temperature: float | None = None
top_p: float | None = None
truncation: str | None = None
user: str | None = None
@json_schema_type
class OpenAIResponseObjectStreamResponseCreated(BaseModel):
response: OpenAIResponseObject
type: Literal["response.created"] = "response.created"
@json_schema_type
class OpenAIResponseObjectStreamResponseCompleted(BaseModel):
response: OpenAIResponseObject
type: Literal["response.completed"] = "response.completed"
OpenAIResponseObjectStream = Annotated[
OpenAIResponseObjectStreamResponseCreated | OpenAIResponseObjectStreamResponseCompleted,
Field(discriminator="type"),
]
register_schema(OpenAIResponseObjectStream, name="OpenAIResponseObjectStream")
@json_schema_type
class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
"""
This represents the output of a function call that gets passed back to the model.
"""
call_id: str
output: str
type: Literal["function_call_output"] = "function_call_output"
id: str | None = None
status: str | None = None
OpenAIResponseInput = Annotated[
# Responses API allows output messages to be passed in as input
OpenAIResponseOutputMessageWebSearchToolCall
| OpenAIResponseOutputMessageFunctionToolCall
| OpenAIResponseInputFunctionToolCallOutput
|
# Fallback to the generic message type as a last resort
OpenAIResponseMessage,
Field(union_mode="left_to_right"),
]
register_schema(OpenAIResponseInput, name="OpenAIResponseInput")
@json_schema_type
class OpenAIResponseInputToolWebSearch(BaseModel):
type: Literal["web_search"] | Literal["web_search_preview_2025_03_11"] = "web_search"
# TODO: actually use search_context_size somewhere...
search_context_size: str | None = Field(default="medium", pattern="^low|medium|high$")
# TODO: add user_location
@json_schema_type
class OpenAIResponseInputToolFunction(BaseModel):
type: Literal["function"] = "function"
name: str
description: str | None = None
parameters: dict[str, Any] | None
strict: bool | None = None
class FileSearchRankingOptions(BaseModel):
ranker: str | None = None
score_threshold: float | None = Field(default=0.0, ge=0.0, le=1.0)
@json_schema_type
class OpenAIResponseInputToolFileSearch(BaseModel):
type: Literal["file_search"] = "file_search"
vector_store_id: list[str]
ranking_options: FileSearchRankingOptions | None = None
# TODO: add filters
OpenAIResponseInputTool = Annotated[
OpenAIResponseInputToolWebSearch | OpenAIResponseInputToolFileSearch | OpenAIResponseInputToolFunction,
Field(discriminator="type"),
]
register_schema(OpenAIResponseInputTool, name="OpenAIResponseInputTool")
class OpenAIResponseInputItemList(BaseModel):
data: list[OpenAIResponseInput]
object: Literal["list"] = "list"