Errors in _ExtractUpdates don't propagate

[mokrueger]

Hi,

i noticed that the errors in the _Extract class don't propagate through the invoke call. This means that if the patch extraction fails, there will not be a retry.

aka. this will not be triggered.

trustcall/trustcall/_base.py

Lines 1005 to 1014 in 8c7312b

	except Exception as e:
	return {
	"messages": [
	HumanMessage(
	content="Fix the validation error while"
	f" also avoiding: {repr(str(e))}"
	)
	],
	"attempts": 1,
	}

This is relevant if the llm does not generate the indices properly and applying the patch fails.

trustcall/trustcall/_base.py

Lines 806 to 825 in 8c7312b

	if target:
	try:
	patches = _ensure_patches(tc["args"])
	if patches or self.tool_choice == "PatchDoc":
	# The second condition is so that, when we are continuously
	# updating a single doc, we will still include it in
	# the output responses list; mainly for backwards
	# compatibility
	resolved_tool_calls.append(
	ToolCall(
	id=tc["id"],
	name=tool_name,
	args=_apply_patch(target, patches),
	)
	)
	updated_docs[tc["id"]] = str(json_doc_id)
	except Exception as e:
	logger.error(f"Could not apply patch: {e}")
	if rt:
	rt.error = f"Could not apply patch: {repr(e)}"

We should re raise the error such that we can retry properly.
If you want I could draft a PR. Any input would be appreciated.