You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
Right now we're limited to either implement tool calling within a function/pipeline directly, or using the default function calling pipeline.
By re-implementing tool calling in a pipeline, we have to define all tools directly, which makes it harder for users to pull extra tools. Additionally, we can't select/de-select them from the UI.
By using native tool calling, we can't control when tools are called (for example, it's impossible to implement an OpenAI-like tool call in the middle of a message), and in the case of functions, we can't know when the pipe() function is called to generate tool call responses.
Describe the solution you'd like
For a function/pipeline, add an optional attribute called "prepend_tools" that decides if the tools are called and appended to the system prompt before responding.
Add an optional parameter to pipe(), called, say, __tool_signatures__ and __tool_callables__ (or simply __tools__: dict) with __tools__["signatures"] and __tools__["callables"] pointing to signatures (in OpenAI JSON format) and and callables respectively
The text was updated successfully, but these errors were encountered:
tjbck
changed the title
Allow functions/pipelines to access interact with tools natively
enh: Allow functions/pipelines to access interact with tools natively
Jul 9, 2024
Is your feature request related to a problem? Please describe.
Right now we're limited to either implement tool calling within a function/pipeline directly, or using the default function calling pipeline.
Describe the solution you'd like
For a function/pipeline, add an optional attribute called "prepend_tools" that decides if the tools are called and appended to the system prompt before responding.
Add an optional parameter to pipe(), called, say,
__tool_signatures__
and__tool_callables__
(or simply__tools__: dict
) with__tools__["signatures"]
and__tools__["callables"]
pointing to signatures (in OpenAI JSON format) and and callables respectivelyThe text was updated successfully, but these errors were encountered: