`HuggingFaceAPIChatGenerator` incorrectly assumes that `arguments` in a prepared Tool Call is always a dict

[anakin87]

Investigating on test_live_run_with_tools, I found out that now arguments in a Tool Call prepared via the Hugging Face APIs can be a string instead of a dict.

While this is inconsistent with their docs, my impression is that in future they will enforce this behavior (code in main).

For this reason, I propose handling the conversion to dict on our side if arguments is a string.

[YassinNouh21]

Hi @anakin87, I looked into it. implemented a fix—just wanted to confirm if this approach aligns with what you had in mind:
• In both _run_non_streaming and _run_non_streaming_async methods of HuggingFaceAPIChatGenerator, I added a check for hfapi_tc.function.arguments.
• If it’s a string, I try to parse it as JSON:
• If parsing succeeds, it’s treated as a proper dict.
• If parsing fails, I wrap the raw string in a dict under the key "raw_arguments" to maintain structure.
• I also updated the ToolCall class comment to reflect this potential inconsistency with Hugging Face’s docs.
• Added tests for both valid and invalid JSON cases, covering sync and async paths.
• Logged this change in the release notes as well.

Let me know if this handling logic makes sense or if you think it should be done differently!

[anakin87]

Hey, thanks!

In general, I more or less agree with you.

If parsing fails: I would simply show a warning similar to what we do for OpenAI:

haystack/haystack/components/generators/chat/openai.py

Lines 546 to 548 in c3bce46

	except json.JSONDecodeError:
	logger.warning(
	"OpenAI returned a malformed JSON string for tool call arguments. This tool call "

Feel free to open a PR and tag me.