Decode weather tool arguments before function feedback

[dajiaohuang]

Problem

In pe-function-calling.ipynb, the final tool-feedback example passes function.arguments directly into get_current_weather(location, unit="fahrenheit").

Those arguments are JSON text, not a location string. For a supported tool call with {"location":"Boston","unit":"celsius"}, the tool result instead contains the entire JSON string as location and silently uses fahrenheit. The notebook's saved messages output already shows the nested-JSON location for its Boston example.

Change

Decode the arguments and unpack them as keyword arguments, matching the earlier args = json.loads(...) / get_current_weather(**args) example in this same notebook and the OpenAI cookbook argument-handling contract. Clear the affected downstream outputs/execution counts rather than present stale results as a new model run. No model/API version changes.

Validation

Offline fixture executes the notebook's actual dummy weather function and tool-feedback cell, with synthetic assistant/tool-call messages:

• Base fails: location is JSON text and the requested Celsius unit is lost.
• Fixed passes: location is Boston, unit is Celsius and the tool-call ID is preserved.
• A location-only call retains the existing Fahrenheit default.
• All code cells compile; a provider fixture verifies the completion wrapper forwards tools/tool_choice and preserves its existing model default.
• git diff --check passes.

No LLM/API calls, environment-key reads or real weather requests were performed. Full live notebook execution is not claimed.

[vercel]

@dajiaohuang is attempting to deploy a commit to the DAIR-AI Team on Vercel.

A member of the Team first needs to authorize it.