[BFCL] Resolve Issue in Gemini Model When No Model Output (#809)

When the Gemini models output no content, the `api_response.candidates`
will be `[]`, and so calling `api_response.candidates[0].content.parts`
will give `IndexError: list index out of range`. We should check if
there is a candidate before accessing its content.

berkeley-function-call-leaderboard/.env.example

@@ -6,7 +6,7 @@ FIREWORKS_API_KEY=
 ANTHROPIC_API_KEY=
 NVIDIA_API_KEY=nvapi-XXXXXX
 YI_API_KEY=
-GROK_API_KEY=
+GROK_API_KEY=xai-XXXXXX
 GOGOAGENT_API_KEY=
 WRITER_API_KEY=
 

berkeley-function-call-leaderboard/CHANGELOG.md

@@ -3,6 +3,7 @@
 All notable changes to the Berkeley Function Calling Leaderboard will be documented in this file.
 
 - [Dec 3, 2024] [#810](https://github.com/ShishirPatil/gorilla/pull/810): Add new model `grok-beta` to the leaderboard.
+- [Dec 2, 2024] [#809](https://github.com/ShishirPatil/gorilla/pull/809): Resolve issue in Gemini model when no model output.
 - [Dec 2, 2024] [#808](https://github.com/ShishirPatil/gorilla/pull/808): Improve latency measurement accuracy.
 - [Nov 26, 2024] [#755](https://github.com/ShishirPatil/gorilla/pull/755): Add new model `palmyra-x-004` to the leaderboard.
 - [Nov 25, 2024] [#718](https://github.com/ShishirPatil/gorilla/pull/718): Add new model `openbmb/MiniCPM3-4B-FC` to the leaderboard.

berkeley-function-call-leaderboard/README.md

@@ -99,7 +99,7 @@ FIREWORKS_API_KEY=
 ANTHROPIC_API_KEY=
 NVIDIA_API_KEY=nvapi-XXXXXX
 YI_API_KEY=
-GROK_API_KEY=
+GROK_API_KEY=xai-XXXXXX
 GOGOAGENT_API_KEY=
 WRITER_API_KEY=
 

berkeley-function-call-leaderboard/bfcl/model_handler/oss_model/salesforce.py

@@ -1,4 +1,5 @@
 import json
+import time
 
 from bfcl.model_handler.model_style import ModelStyle
 from bfcl.model_handler.oss_model.base_oss_handler import OSSHandler
@@ -114,9 +115,11 @@ def _query_prompting(self, inference_data: dict):
             "function": function,
         }
 
+        start_time = time.time()
         api_response = self.client.completion(messages=message, tools=function)
+        end_time = time.time()
 
-        return api_response
+        return api_response, end_time - start_time
 
 
 # fmt: off

berkeley-function-call-leaderboard/bfcl/model_handler/proprietary_model/gemini.py

@@ -123,7 +123,7 @@ def _query_FC(self, inference_data: dict):
             generation_config=GenerationConfig(
                 temperature=self.temperature,
             ),
-            tools=tools
+            tools=tools,
         )
 
     def _pre_query_processing_FC(self, inference_data: dict, test_entry: dict) -> dict:
@@ -152,33 +152,38 @@ def _compile_tools(self, inference_data: dict, test_entry: dict) -> dict:
         return inference_data
 
     def _parse_query_response_FC(self, api_response: any) -> dict:
-        response_function_call_content = api_response.candidates[0].content
         tool_call_func_names = []
         fc_parts = []
         text_parts = []
-        for part in api_response.candidates[0].content.parts:
-            # part.function_call is a FunctionCall object, so it will always be True even if it contains no function call
-            # So we need to check if the function name is empty `""` to determine if Gemini returned a function call
-            if part.function_call and part.function_call.name:
-                part_func_name = part.function_call.name
-                part_func_args = part.function_call.args
-                part_func_args_dict = {k: v for k, v in part_func_args.items()}
-
-                fc_parts.append({part_func_name: part_func_args_dict})
-                tool_call_func_names.append(part_func_name)
-            else:
-                text_parts.append(part.text)
 
-        model_responses = fc_parts if fc_parts else text_parts
-
-        if len(api_response.candidates[0].content.parts) == 0:
+        if (
+            len(api_response.candidates) > 0
+            and len(api_response.candidates[0].content.parts) > 0
+        ):
+            response_function_call_content = api_response.candidates[0].content
+
+            for part in api_response.candidates[0].content.parts:
+                # part.function_call is a FunctionCall object, so it will always be True even if it contains no function call
+                # So we need to check if the function name is empty `""` to determine if Gemini returned a function call
+                if part.function_call and part.function_call.name:
+                    part_func_name = part.function_call.name
+                    part_func_args = part.function_call.args
+                    part_func_args_dict = {k: v for k, v in part_func_args.items()}
+
+                    fc_parts.append({part_func_name: part_func_args_dict})
+                    tool_call_func_names.append(part_func_name)
+                else:
+                    text_parts.append(part.text)
+        else:
             response_function_call_content = Content(
                 role="model",
                 parts=[
                     Part.from_text("The model did not return any response."),
                 ],
             )
 
+        model_responses = fc_parts if fc_parts else text_parts
+
         return {
             "model_responses": model_responses,
             "model_responses_message_for_chat_history": response_function_call_content,
@@ -288,7 +293,10 @@ def _pre_query_processing_prompting(self, test_entry: dict) -> dict:
             return {"message": []}
 
     def _parse_query_response_prompting(self, api_response: any) -> dict:
-        if len(api_response.candidates[0].content.parts) > 0:
+        if (
+            len(api_response.candidates) > 0
+            and len(api_response.candidates[0].content.parts) > 0
+        ):
             model_responses = api_response.text
         else:
             model_responses = "The model did not return any response."