-
Notifications
You must be signed in to change notification settings - Fork 1k
Add builtin_tools
to Agent
#1722
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
for tool in builtin_tools: | ||
if tool == 'web-search': | ||
self._builtin_tools.append(WebSearchTool()) | ||
else: | ||
self._builtin_tools.append(tool) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's easier to not have to handle string on the models, so we already do the transformation here.
""" | ||
|
||
|
||
class UserLocation(TypedDict, total=False): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's easier to handle this in the models if it's a TypedDict
, since it matches the type.
@dataclass | ||
class WebSearchTool(AbstractBuiltinTool): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
From the DX POV, it's nicer for it to be a BaseModel
or dataclass
.
Docs Preview
|
tools = list(model_settings.get('openai_builtin_tools', [])) + tools | ||
tools = self._get_builtin_tools(model_request_parameters) + tools |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We should deprecate the openai_builtin_tools
in this PR.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So we'll add support for FileSearchToolParam and ComputerToolParam as well?
Note that https://platform.openai.com/docs/guides/tools-file-search also results in a "type": "file_search_call"
output item.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do we still want to allow people to use arbitrary built-in tools we haven't created a class for yet by passing the appropriate JSON?
class AbstractBuiltinTool(ABC): | ||
"""A builtin tool that can be used by an agent. | ||
|
||
This class is abstract and cannot be instantiated directly. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think worth including a sentence here explaining how the code execution works to make use of them — something like "these are passed to the model as part of the ModelRequestParameters" or whatever. (Not sure if that's true, haven't gotten there yet ..). But I imagine it helping someone who is trying to figure out how they are different from normal tools.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM other than tests and docs
Co-authored-by: Marcelo Trylesinski <marcelotryle@gmail.com>
PR Change SummaryEnhanced documentation for the
Modified Files
How can I customize these reviews?Check out the Hyperlint AI Reviewer docs for more information on how to customize the review. If you just want to ignore it on this PR, you can add the Note specifically for link checks, we only check the first 30 links in a file and we cache the results for several hours (for instance, if you just added a page, you might experience this). Our recommendation is to add |
https://github.com/mattbrandman/pydantic-ai/pull/1/files Working on a fork that incorporates the new responses api code interpreter and handles streaming. I'm sure code quality is questionable but feel free to incorporate anything thats useful! |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
tools = list(model_settings.get('openai_builtin_tools', [])) + tools | ||
tools = self._get_builtin_tools(model_request_parameters) + tools |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So we'll add support for FileSearchToolParam and ComputerToolParam as well?
Note that https://platform.openai.com/docs/guides/tools-file-search also results in a "type": "file_search_call"
output item.
user_location=user_location, | ||
) | ||
) | ||
elif isinstance(tool, CodeExecutionTool): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think we'll want an else
-> raise "Unsupported built-in tool"
here as well (and in OpenAI and Groq, looks like only Google has it currently)
def _get_builtin_tools(self, model_request_parameters: ModelRequestParameters) -> list[responses.ToolParam]: | ||
tools: list[responses.ToolParam] = [] | ||
for tool in model_request_parameters.builtin_tools: | ||
if isinstance(tool, WebSearchTool): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We still need to implement CodeExecutionTool here, right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The linked commit I have has a basic implementation. Not sure if it would be best to open a PR into this branch or add it after this is merged. There’s unfortunately a bunch of type errors (or were last I checked) from OpenAI where required fields are empty so need a series of guards that look redundant to checkers
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
No, chat completions doesn't have it.
tools = list(model_settings.get('openai_builtin_tools', [])) + tools | ||
tools = self._get_builtin_tools(model_request_parameters) + tools |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do we still want to allow people to use arbitrary built-in tools we haven't created a class for yet by passing the appropriate JSON?
@@ -694,6 +721,20 @@ def _get_tools(self, model_request_parameters: ModelRequestParameters) -> list[r | |||
tools += [self._map_tool_definition(r) for r in model_request_parameters.output_tools] | |||
return tools | |||
|
|||
def _get_builtin_tools(self, model_request_parameters: ModelRequestParameters) -> list[responses.ToolParam]: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The case I was thinking of that I mentioned on the call this morning is that someone may use OpenAIModel
with OpenRouterProvider
and an Anthropic model, and it's possible that the built-in tool specifications should then match those for Anthropic rather than those for OpenAI. But from a quick search it doesn't look like OpenRouter currently supports built-in tools (at least there's no doc on it), and when they do they'll probably require a OR-specific specification that they'll translate to the model-specific specs behind the scenes, so we don't have to worry about that for now.
besides having tests pass is there anything else holding back this PR seems like it will grow in importance as more builtin tools are introduced? |
@Kludex not sure if this will be helpful but have a branch based off this with merge conflicts handled and tests running properly. Don't know if I did everything right and also think I should probably have a cleaner one that could be merged into this but my git-foo is failing me late at night will see if I can make a cleaner version mattbrandman#6 |
@mattbrandman Thanks for creating the #2102 PR! @Kludex is out this week but I'm sure he'll be happy to see the work you've saved him when he's back :) |
@DouweM no problem! I'm happy to have them merged back into here just couldn't figure out a clean PR strategy since I had to merge main to resolve the conflicts. |
No description provided.