fix: fallback to json_object when json_schema response_format is unsupported

src/uipath/_cli/_evals/mocks/llm_mocker.py

@@ -10,6 +10,7 @@
     LLMMockingStrategy,
     MockingContext,
 )
+from uipath.platform.errors import EnrichedException
 from uipath.tracing import traced
 from uipath.tracing._utils import _SpanUtils
 
@@ -92,7 +93,6 @@ async def response(
         function_name = params.get("name") or func.__name__
         if function_name in [x.name for x in self.context.strategy.tools_to_simulate]:
             from uipath.platform import UiPath
-            from uipath.platform.chat._llm_gateway_service import _cleanup_schema
 
             from .mocks import (
                 cache_manager_context,
@@ -110,6 +110,8 @@ async def response(
             if return_type is None:
                 return_type = Any
 
+            from uipath.platform.chat._llm_gateway_service import _cleanup_schema
+
             output_schema = params.get(
                 "output_schema", TypeAdapter(return_type).json_schema()
             )
@@ -195,16 +197,33 @@ async def response(
                     if cached_response is not None:
                         return cached_response
 
-                response = await llm.chat_completions(
-                    [
-                        {
-                            "role": "user",
-                            "content": formatted_prompt,
-                        },
-                    ],
-                    response_format=response_format,
-                    **completion_kwargs,
-                )
+                messages = [
+                    {
+                        "role": "user",
+                        "content": formatted_prompt,
+                    },
+                ]
+
+                try:
+                    response = await llm.chat_completions(
+                        messages,
+                        response_format=response_format,
+                        **completion_kwargs,
+                    )
+                except EnrichedException as e:
+                    if e.status_code in (400, 422):
+                        logger.warning(
+                            "json_schema response_format not supported by model, "
+                            "falling back to json_object"
+                        )
+                        response = await llm.chat_completions(
+                            messages,
+                            response_format={"type": "json_object"},
+                            **completion_kwargs,
+                        )
+                    else:
+                        raise
+
                 result = json.loads(response.choices[0].message.content)
 
                 if cache_manager is not None: