cleaning final answer

Changelog.md

@@ -12,6 +12,7 @@
 - Simplified error handling in `run_and_submit_all` by removing redundant try-except block around `instantiate_agent` call.
 - Updated `run_and_submit_all` to call `fetch_questions` and handle its return values.
 - Updated `app_dev.py` to import and use `SmolAgent` from `my_agent.py`.
+- Added logic to `GeminiApiModel` to extract text after "FINAL ANSWER: " marker in responses.
 
 ### Deprecated
 

gemini_model.py

@@ -125,6 +125,11 @@ class GeminiApiModel(ApiModel):
                 part = response_json["candidates"][0]["content"]["parts"][0]
                 if "text" in part:
                     content = part["text"]
+                    # Check for "FINAL ANSWER: " and extract the rest of the string
+                    final_answer_marker = "FINAL ANSWER: "
+                    if final_answer_marker in content:
+                        content = content.split(final_answer_marker)[-1].strip()
+
                     # Simulate token counts if available, otherwise default to 0
                     # The basic generateContent API might not return usage directly in the main response
                     # It might be in safetyRatings or other metadata if enabled/available.