fix: parse empty tool call arguments in the non-streaming OpenAI generators

haystack/components/generators/chat/openai.py

@@ -39,6 +39,7 @@
     ToolCallDelta,
     select_streaming_callback,
 )
+from haystack.dataclasses.chat_message import _parse_openai_tool_call_arguments
 from haystack.dataclasses.streaming_chunk import _invoke_streaming_callback
 from haystack.tools import (
     ToolsType,
@@ -668,7 +669,7 @@ def _convert_chat_completion_to_chat_message(
         for openai_tc in openai_tool_calls:
             arguments_str = openai_tc.function.arguments
             try:
-                arguments = json.loads(arguments_str)
+                arguments = _parse_openai_tool_call_arguments(arguments_str)
                 tool_calls.append(ToolCall(id=openai_tc.id, tool_name=openai_tc.function.name, arguments=arguments))
             except json.JSONDecodeError:
                 logger.warning(

haystack/components/generators/chat/openai_responses.py

@@ -29,6 +29,7 @@
     ToolCallDelta,
     select_streaming_callback,
 )
+from haystack.dataclasses.chat_message import _parse_openai_tool_call_arguments
 from haystack.dataclasses.streaming_chunk import FinishReason, _invoke_streaming_callback
 from haystack.tools import (
     ToolsType,
@@ -692,7 +693,7 @@ def _convert_response_to_chat_message(responses: Response | ParsedResponse) -> C
 
         elif output.type == "function_call":
             try:
-                arguments = json.loads(output.arguments)
+                arguments = _parse_openai_tool_call_arguments(output.arguments)
                 tool_calls.append(
                     ToolCall(
                         id=output.id, tool_name=output.name, arguments=arguments, extra={"call_id": output.call_id}

releasenotes/notes/fix-openai-zero-argument-tool-call-e616b87f179be88b.yaml

@@ -0,0 +1,7 @@
+---
+fixes:
+  - |
+    ``OpenAIChatGenerator`` and ``OpenAIResponsesChatGenerator`` no longer drop a tool call whose arguments are an
+    empty string, and no longer raise ``TypeError`` when they are ``null``, in non-streaming responses.
+    OpenAI-compatible servers such as vLLM send these for a tool with no parameters. Both now become empty
+    arguments, as they already did in streaming responses and in ``ChatMessage.from_openai_dict_format``.

test/components/generators/chat/test_openai.py

@@ -37,6 +37,7 @@
     OpenAIChatGenerator,
     _check_finish_reason,
     _convert_chat_completion_chunk_to_streaming_chunk,
+    _convert_chat_completion_to_chat_message,
     _make_schema_strict,
 )
 from haystack.components.generators.utils import print_streaming_chunk
@@ -2031,6 +2032,34 @@ def test_convert_usage_chunk_to_streaming_chunk(self) -> None:
         assert result.meta["model"] == "gpt-5-mini"
         assert result.meta["received_at"] is not None
 
+    @pytest.mark.parametrize("arguments", ["", None])
+    def test_convert_chat_completion_with_zero_argument_tool_call(self, arguments: str | None) -> None:
+        # OpenAI-compatible servers such as vLLM send an empty string or null for a tool with no parameters
+        completion = ChatCompletion(
+            id="1",
+            model="gpt-5-mini",
+            object="chat.completion",
+            created=1234567890,
+            choices=[
+                Choice(
+                    finish_reason="tool_calls",
+                    index=0,
+                    message=ChatCompletionMessage(
+                        role="assistant",
+                        tool_calls=[
+                            ChatCompletionMessageFunctionToolCall(
+                                id="1",
+                                type="function",
+                                function=Function.model_construct(name="get_time", arguments=arguments),
+                            )
+                        ],
+                    ),
+                )
+            ],
+        )
+        message = _convert_chat_completion_to_chat_message(completion, completion.choices[0])
+        assert message.tool_calls == [ToolCall(id="1", tool_name="get_time", arguments={})]
+
 
 class TestMakeSchemaStrict:
     def test_flat_object(self) -> None:

test/components/generators/chat/test_openai_responses_conversion.py

@@ -1296,6 +1296,20 @@ def test_convert_incomplete_response_with_finish_reason(
 
         assert message.meta["finish_reason"] == finish_reason
 
+    @pytest.mark.parametrize("arguments", ["", None])
+    def test_convert_zero_argument_function_call(self, arguments: str | None) -> None:
+        # OpenAI-compatible servers such as vLLM send an empty string or null for a tool with no parameters
+        function_call = ResponseFunctionToolCall.model_construct(
+            arguments=arguments, call_id="call_1", name="get_time", type="function_call", id="fc_1", status="completed"
+        )
+        response = Response.model_construct(output=[function_call], output_text=None, status="completed")
+
+        message = _convert_response_to_chat_message(response)
+
+        assert message.tool_calls == [
+            ToolCall(id="fc_1", tool_name="get_time", arguments={}, extra={"call_id": "call_1"})
+        ]
+
     def test_convert_system_message(self) -> None:
 
         message = ChatMessage.from_system("You are good assistant")