Improve quality of ollama tool calling by repairing arguments (#122749)

* Improve quality of ollama function calling by repairing function call arguments

* Fix formatting of the tests

* Run ruff format on ollama conversation

* Add test for non-string arguments

homeassistant/components/ollama/conversation.py

@@ -63,6 +63,34 @@ def _format_tool(
     return {"type": "function", "function": tool_spec}
 
 
+def _fix_invalid_arguments(value: Any) -> Any:
+    """Attempt to repair incorrectly formatted json function arguments.
+
+    Small models (for example llama3.1 8B) may produce invalid argument values
+    which we attempt to repair here.
+    """
+    if not isinstance(value, str):
+        return value
+    if (value.startswith("[") and value.endswith("]")) or (
+        value.startswith("{") and value.endswith("}")
+    ):
+        try:
+            return json.loads(value)
+        except json.decoder.JSONDecodeError:
+            pass
+    return value
+
+
+def _parse_tool_args(arguments: dict[str, Any]) -> dict[str, Any]:
+    """Rewrite ollama tool arguments.
+
+    This function improves tool use quality by fixing common mistakes made by
+    small local tool use models. This will repair invalid json arguments and
+    omit unnecessary arguments with empty values that will fail intent parsing.
+    """
+    return {k: _fix_invalid_arguments(v) for k, v in arguments.items() if v}
+
+
 class OllamaConversationEntity(
     conversation.ConversationEntity, conversation.AbstractConversationAgent
 ):
@@ -255,7 +283,7 @@ class OllamaConversationEntity(
             for tool_call in tool_calls:
                 tool_input = llm.ToolInput(
                     tool_name=tool_call["function"]["name"],
-                    tool_args=tool_call["function"]["arguments"],
+                    tool_args=_parse_tool_args(tool_call["function"]["arguments"]),
                 )
                 _LOGGER.debug(
                     "Tool call: %s(%s)", tool_input.tool_name, tool_input.tool_args

tests/components/ollama/test_conversation.py

@@ -1,5 +1,6 @@
 """Tests for the Ollama integration."""
 
+from typing import Any
 from unittest.mock import AsyncMock, Mock, patch
 
 from ollama import Message, ResponseError
@@ -116,12 +117,33 @@ async def test_template_variables(
     assert "The user id is 12345." in prompt
 
 
+@pytest.mark.parametrize(
+    ("tool_args", "expected_tool_args"),
+    [
+        ({"param1": "test_value"}, {"param1": "test_value"}),
+        ({"param1": 2}, {"param1": 2}),
+        (
+            {"param1": "test_value", "floor": ""},
+            {"param1": "test_value"},  # Omit empty arguments
+        ),
+        (
+            {"domain": '["light"]'},
+            {"domain": ["light"]},  # Repair invalid json arguments
+        ),
+        (
+            {"domain": "['light']"},
+            {"domain": "['light']"},  # Preserve invalid json that can't be parsed
+        ),
+    ],
+)
 @patch("homeassistant.components.ollama.conversation.llm.AssistAPI._async_get_tools")
 async def test_function_call(
     mock_get_tools,
     hass: HomeAssistant,
     mock_config_entry_with_assist: MockConfigEntry,
     mock_init_component,
+    tool_args: dict[str, Any],
+    expected_tool_args: dict[str, Any],
 ) -> None:
     """Test function call from the assistant."""
     agent_id = mock_config_entry_with_assist.entry_id
@@ -154,7 +176,7 @@ async def test_function_call(
                     {
                         "function": {
                             "name": "test_tool",
-                            "arguments": {"param1": "test_value"},
+                            "arguments": tool_args,
                         }
                     }
                 ],
@@ -183,7 +205,7 @@ async def test_function_call(
         hass,
         llm.ToolInput(
             tool_name="test_tool",
-            tool_args={"param1": "test_value"},
+            tool_args=expected_tool_args,
         ),
         llm.LLMContext(
             platform="ollama",