Update app.py
Browse files
app.py
CHANGED
|
@@ -35,28 +35,13 @@ class BasicAgent:
|
|
| 35 |
Process a GAIA benchmark question and return the answer
|
| 36 |
"""
|
| 37 |
try:
|
| 38 |
-
# Ensure question is a string
|
| 39 |
-
if not isinstance(question, str):
|
| 40 |
-
question = str(question)
|
| 41 |
-
|
| 42 |
if self.verbose:
|
| 43 |
print(f"Processing question: {question}")
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
# CRITICAL FIX: Ensure task_file_path is a string or None
|
| 47 |
-
if task_file_path is not None:
|
| 48 |
-
if isinstance(task_file_path, list):
|
| 49 |
-
task_file_path = task_file_path[0] if len(task_file_path) > 0 else None
|
| 50 |
-
print(f"Converted list to string: {task_file_path}")
|
| 51 |
-
elif not isinstance(task_file_path, str):
|
| 52 |
-
task_file_path = str(task_file_path)
|
| 53 |
-
print(f"Converted to string: {task_file_path}")
|
| 54 |
-
|
| 55 |
-
if self.verbose and task_file_path:
|
| 56 |
-
print(f"With associated file: {task_file_path}")
|
| 57 |
|
| 58 |
# Create a context with file information if available
|
| 59 |
-
context =
|
| 60 |
|
| 61 |
# If there's a file, read it and include its content in the context
|
| 62 |
if task_file_path:
|
|
@@ -77,10 +62,10 @@ You can still try to answer the question based on the information provided.
|
|
| 77 |
"""
|
| 78 |
|
| 79 |
# Check for special cases that need specific formatting
|
| 80 |
-
if
|
| 81 |
context = f"""
|
| 82 |
This question appears to be in reversed text. Here's the reversed version:
|
| 83 |
-
{
|
| 84 |
Now answer the question above. Remember to format your answer exactly as requested.
|
| 85 |
"""
|
| 86 |
|
|
@@ -92,48 +77,16 @@ Be direct and specific. GAIA benchmark requires exact matching answers.
|
|
| 92 |
For example, if asked "What is the capital of France?", respond simply with "Paris".
|
| 93 |
"""
|
| 94 |
|
| 95 |
-
#
|
| 96 |
-
full_prompt = str(full_prompt)
|
| 97 |
-
|
| 98 |
-
if self.verbose:
|
| 99 |
-
print(f"Full prompt type: {type(full_prompt)}")
|
| 100 |
-
print(f"Full prompt length: {len(full_prompt)}")
|
| 101 |
-
|
| 102 |
-
# *** FIXED AGENT CALL - Pass string to CodeAgent ***
|
| 103 |
try:
|
| 104 |
-
# CodeAgent.run expects a textual task (string), not a list of messages
|
| 105 |
-
# Passing a list causes internal `.strip()` calls on a list, which raises:
|
| 106 |
-
# "AttributeError: 'list' object has no attribute 'strip'"
|
| 107 |
raw_response = self.agent.run(full_prompt)
|
| 108 |
|
| 109 |
if self.verbose:
|
| 110 |
print(f"Raw response type: {type(raw_response)}")
|
| 111 |
print(f"Raw response: {raw_response}")
|
| 112 |
|
| 113 |
-
# Handle ALL possible response formats
|
| 114 |
-
|
| 115 |
-
answer = raw_response.get('choices', [{}])[0].get('message', {}).get('content', str(raw_response))
|
| 116 |
-
elif isinstance(raw_response, list):
|
| 117 |
-
if len(raw_response) > 0:
|
| 118 |
-
if isinstance(raw_response[0], dict):
|
| 119 |
-
# Common format: [{"role": "assistant", "content": "..."}]
|
| 120 |
-
answer = raw_response[0].get('content', str(raw_response[0]))
|
| 121 |
-
elif isinstance(raw_response[0], list):
|
| 122 |
-
# Nested list - dig deeper
|
| 123 |
-
nested = raw_response[0]
|
| 124 |
-
if isinstance(nested, list) and len(nested) > 0:
|
| 125 |
-
if isinstance(nested[0], dict):
|
| 126 |
-
answer = nested[0].get('content', str(nested[0]))
|
| 127 |
-
else:
|
| 128 |
-
answer = str(nested[0])
|
| 129 |
-
else:
|
| 130 |
-
answer = str(raw_response[0])
|
| 131 |
-
else:
|
| 132 |
-
answer = str(raw_response[0])
|
| 133 |
-
else:
|
| 134 |
-
answer = "No response"
|
| 135 |
-
else:
|
| 136 |
-
answer = str(raw_response)
|
| 137 |
|
| 138 |
if self.verbose:
|
| 139 |
print(f"Extracted answer type: {type(answer)}")
|
|
@@ -145,13 +98,7 @@ For example, if asked "What is the capital of France?", respond simply with "Par
|
|
| 145 |
return f"Agent error: {agent_error}"
|
| 146 |
|
| 147 |
# Clean the answer
|
| 148 |
-
|
| 149 |
-
answer = self._clean_answer(answer)
|
| 150 |
-
except Exception as clean_error:
|
| 151 |
-
print(f"Error cleaning answer: {clean_error}")
|
| 152 |
-
traceback.print_exc()
|
| 153 |
-
# Just return the raw answer as string if cleaning fails
|
| 154 |
-
answer = str(answer) if answer else "Error cleaning answer"
|
| 155 |
|
| 156 |
if self.verbose:
|
| 157 |
print(f"Generated answer: {answer}")
|
|
@@ -165,28 +112,71 @@ For example, if asked "What is the capital of France?", respond simply with "Par
|
|
| 165 |
traceback.print_exc()
|
| 166 |
return error_msg
|
| 167 |
|
| 168 |
-
def
|
| 169 |
"""
|
| 170 |
-
|
|
|
|
| 171 |
"""
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
return ""
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
if not isinstance(answer, str):
|
| 186 |
answer = str(answer)
|
|
|
|
| 187 |
except Exception as e:
|
| 188 |
print(f"Error in initial conversion: {e}")
|
| 189 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 190 |
|
| 191 |
# Now answer should definitely be a string
|
| 192 |
try:
|
|
@@ -211,23 +201,21 @@ For example, if asked "What is the capital of France?", respond simply with "Par
|
|
| 211 |
|
| 212 |
return answer
|
| 213 |
except Exception as e:
|
| 214 |
-
print(f"Error in answer cleaning: {e}
|
| 215 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
|
| 217 |
def run_and_submit_all(profile: gr.OAuthProfile | None):
|
| 218 |
"""
|
| 219 |
Fetches all questions, runs the BasicAgent on them, submits all answers,
|
| 220 |
and displays the results.
|
| 221 |
"""
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
username = f"{profile.username}"
|
| 226 |
-
print(f"User logged in: {username}")
|
| 227 |
-
else:
|
| 228 |
-
print("User not logged in.")
|
| 229 |
-
return "Please Login to Hugging Face with the button.", None
|
| 230 |
-
|
| 231 |
api_url = DEFAULT_API_URL
|
| 232 |
questions_url = f"{api_url}/questions"
|
| 233 |
submit_url = f"{api_url}/submit"
|
|
@@ -286,6 +274,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 286 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
| 287 |
except Exception as e:
|
| 288 |
print(f"Error running agent on task {task_id}: {e}")
|
|
|
|
| 289 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
| 290 |
|
| 291 |
if not answers_payload:
|
|
@@ -337,6 +326,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 337 |
except Exception as e:
|
| 338 |
status_message = f"An unexpected error occurred during submission: {e}"
|
| 339 |
print(status_message)
|
|
|
|
| 340 |
results_df = pd.DataFrame(results_log)
|
| 341 |
return status_message, results_df
|
| 342 |
|
|
@@ -392,7 +382,7 @@ class Tee:
|
|
| 392 |
|
| 393 |
if __name__ == "__main__":
|
| 394 |
# Redirect stdout and stderr
|
| 395 |
-
log_timestamp = datetime.now().strftime("%Y-%m-%
|
| 396 |
log_file = f"./logs/output_{log_timestamp}.log"
|
| 397 |
tee = Tee(log_file)
|
| 398 |
sys.stdout = tee
|
|
|
|
| 35 |
Process a GAIA benchmark question and return the answer
|
| 36 |
"""
|
| 37 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
if self.verbose:
|
| 39 |
print(f"Processing question: {question}")
|
| 40 |
+
if task_file_path:
|
| 41 |
+
print(f"With associated file: {task_file_path}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
# Create a context with file information if available
|
| 44 |
+
context = question
|
| 45 |
|
| 46 |
# If there's a file, read it and include its content in the context
|
| 47 |
if task_file_path:
|
|
|
|
| 62 |
"""
|
| 63 |
|
| 64 |
# Check for special cases that need specific formatting
|
| 65 |
+
if question.startswith(".") or ".rewsna eht sa" in question:
|
| 66 |
context = f"""
|
| 67 |
This question appears to be in reversed text. Here's the reversed version:
|
| 68 |
+
{question[::-1]}
|
| 69 |
Now answer the question above. Remember to format your answer exactly as requested.
|
| 70 |
"""
|
| 71 |
|
|
|
|
| 77 |
For example, if asked "What is the capital of France?", respond simply with "Paris".
|
| 78 |
"""
|
| 79 |
|
| 80 |
+
# *** FIXED AGENT CALL - Handles all response formats ***
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
try:
|
|
|
|
|
|
|
|
|
|
| 82 |
raw_response = self.agent.run(full_prompt)
|
| 83 |
|
| 84 |
if self.verbose:
|
| 85 |
print(f"Raw response type: {type(raw_response)}")
|
| 86 |
print(f"Raw response: {raw_response}")
|
| 87 |
|
| 88 |
+
# Handle ALL possible response formats with recursive flattening
|
| 89 |
+
answer = self._extract_answer(raw_response)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
|
| 91 |
if self.verbose:
|
| 92 |
print(f"Extracted answer type: {type(answer)}")
|
|
|
|
| 98 |
return f"Agent error: {agent_error}"
|
| 99 |
|
| 100 |
# Clean the answer
|
| 101 |
+
answer = self._clean_answer(answer)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
if self.verbose:
|
| 104 |
print(f"Generated answer: {answer}")
|
|
|
|
| 112 |
traceback.print_exc()
|
| 113 |
return error_msg
|
| 114 |
|
| 115 |
+
def _extract_answer(self, response):
|
| 116 |
"""
|
| 117 |
+
Recursively extract the answer from any nested structure.
|
| 118 |
+
This handles dict, list, and nested combinations.
|
| 119 |
"""
|
| 120 |
+
if response is None:
|
| 121 |
+
return ""
|
| 122 |
+
|
| 123 |
+
# If it's already a string, return it
|
| 124 |
+
if isinstance(response, str):
|
| 125 |
+
return response
|
| 126 |
+
|
| 127 |
+
# If it's a dict, try to get 'content' or 'message'
|
| 128 |
+
if isinstance(response, dict):
|
| 129 |
+
# Try common keys
|
| 130 |
+
for key in ['content', 'message', 'text', 'answer']:
|
| 131 |
+
if key in response:
|
| 132 |
+
return self._extract_answer(response[key])
|
| 133 |
+
# Try OpenAI-style format
|
| 134 |
+
if 'choices' in response:
|
| 135 |
+
choices = response['choices']
|
| 136 |
+
if choices and len(choices) > 0:
|
| 137 |
+
return self._extract_answer(choices[0])
|
| 138 |
+
# If nothing found, convert to string
|
| 139 |
+
return str(response)
|
| 140 |
+
|
| 141 |
+
# If it's a list, recursively process elements
|
| 142 |
+
if isinstance(response, list):
|
| 143 |
+
if len(response) == 0:
|
| 144 |
return ""
|
| 145 |
+
# Try to extract from first element
|
| 146 |
+
first_element = self._extract_answer(response[0])
|
| 147 |
+
# If first element is empty or None, try other elements
|
| 148 |
+
if not first_element and len(response) > 1:
|
| 149 |
+
for element in response[1:]:
|
| 150 |
+
extracted = self._extract_answer(element)
|
| 151 |
+
if extracted:
|
| 152 |
+
return extracted
|
| 153 |
+
return first_element
|
| 154 |
+
|
| 155 |
+
# For any other type, convert to string
|
| 156 |
+
return str(response)
|
| 157 |
+
|
| 158 |
+
def _clean_answer(self, answer) -> str:
|
| 159 |
+
"""
|
| 160 |
+
Ultra-safe answer extraction and cleaning.
|
| 161 |
+
Ensures the answer is always a string.
|
| 162 |
+
"""
|
| 163 |
+
# First, ensure we have a string
|
| 164 |
+
try:
|
| 165 |
+
# Recursively extract if needed
|
| 166 |
+
if not isinstance(answer, str):
|
| 167 |
+
answer = self._extract_answer(answer)
|
| 168 |
+
|
| 169 |
+
# At this point, answer should be a string
|
| 170 |
if not isinstance(answer, str):
|
| 171 |
answer = str(answer)
|
| 172 |
+
|
| 173 |
except Exception as e:
|
| 174 |
print(f"Error in initial conversion: {e}")
|
| 175 |
+
traceback.print_exc()
|
| 176 |
+
try:
|
| 177 |
+
return str(answer) if answer else ""
|
| 178 |
+
except:
|
| 179 |
+
return ""
|
| 180 |
|
| 181 |
# Now answer should definitely be a string
|
| 182 |
try:
|
|
|
|
| 201 |
|
| 202 |
return answer
|
| 203 |
except Exception as e:
|
| 204 |
+
print(f"Error in answer cleaning: {e}")
|
| 205 |
+
traceback.print_exc()
|
| 206 |
+
try:
|
| 207 |
+
return str(answer) if answer else ""
|
| 208 |
+
except:
|
| 209 |
+
return ""
|
| 210 |
|
| 211 |
def run_and_submit_all(profile: gr.OAuthProfile | None):
|
| 212 |
"""
|
| 213 |
Fetches all questions, runs the BasicAgent on them, submits all answers,
|
| 214 |
and displays the results.
|
| 215 |
"""
|
| 216 |
+
if profile is None:
|
| 217 |
+
return "Please log in with your Hugging Face account first.", None
|
| 218 |
+
username = profile.username
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
api_url = DEFAULT_API_URL
|
| 220 |
questions_url = f"{api_url}/questions"
|
| 221 |
submit_url = f"{api_url}/submit"
|
|
|
|
| 274 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
| 275 |
except Exception as e:
|
| 276 |
print(f"Error running agent on task {task_id}: {e}")
|
| 277 |
+
traceback.print_exc()
|
| 278 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
| 279 |
|
| 280 |
if not answers_payload:
|
|
|
|
| 326 |
except Exception as e:
|
| 327 |
status_message = f"An unexpected error occurred during submission: {e}"
|
| 328 |
print(status_message)
|
| 329 |
+
traceback.print_exc()
|
| 330 |
results_df = pd.DataFrame(results_log)
|
| 331 |
return status_message, results_df
|
| 332 |
|
|
|
|
| 382 |
|
| 383 |
if __name__ == "__main__":
|
| 384 |
# Redirect stdout and stderr
|
| 385 |
+
log_timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
|
| 386 |
log_file = f"./logs/output_{log_timestamp}.log"
|
| 387 |
tee = Tee(log_file)
|
| 388 |
sys.stdout = tee
|