Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,459 +1,327 @@
|
|
| 1 |
import os
|
| 2 |
-
import
|
| 3 |
-
import requests
|
| 4 |
-
import inspect
|
| 5 |
-
import pandas as pd
|
| 6 |
-
import asyncio
|
| 7 |
-
import aiohttp
|
| 8 |
-
import time
|
| 9 |
-
import random
|
| 10 |
-
import json
|
| 11 |
-
import re
|
| 12 |
-
from smolagents import FinalAnswerTool, Tool, tool, OpenAIServerModel, DuckDuckGoSearchTool, CodeAgent, VisitWebpageTool
|
| 13 |
-
from gemini_agent import GeminiAgent # Assuming you have a GeminiAgent class defined in gemini_agent.py
|
| 14 |
-
|
| 15 |
from dotenv import load_dotenv
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
load_dotenv()
|
| 18 |
-
# (Keep Constants as is)
|
| 19 |
-
# --- Constants ---
|
| 20 |
-
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 21 |
|
| 22 |
-
|
| 23 |
-
OPENAI_TOKEN = os.getenv("OPENAI_API_KEY")
|
| 24 |
-
|
| 25 |
-
# --- Custom Tools for Better Reasoning ---
|
| 26 |
-
|
| 27 |
-
class TrickQuestionDetector(Tool):
|
| 28 |
-
"""Detects and handles trick questions"""
|
| 29 |
-
|
| 30 |
def __init__(self):
|
| 31 |
-
|
| 32 |
-
self.name = "trick_detector"
|
| 33 |
-
self.description = "Analyze if a question is a trick question and provide guidance"
|
| 34 |
-
self.inputs = {"question": {"type": "string", "description": "The question to analyze"}}
|
| 35 |
-
|
| 36 |
-
def detect_trick(self, question: str) -> str:
|
| 37 |
-
"""Detect common trick question patterns"""
|
| 38 |
-
q_lower = question.lower()
|
| 39 |
|
| 40 |
-
#
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
# Simple heuristic: if reversed version has common English patterns
|
| 44 |
-
if any(word in reversed_q.lower() for word in ['the', 'and', 'what', 'how', 'when', 'where']):
|
| 45 |
-
return f"TRICK DETECTED: This appears to be reversed text. Decoded: '{reversed_q}'"
|
| 46 |
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
|
| 51 |
-
#
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
|
|
|
| 55 |
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
return "TRICK DETECTED: Mathematical trick involving zero or impossible calculations."
|
| 59 |
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
|
| 71 |
-
def
|
| 72 |
-
"""
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
if any(word in q_lower for word in ['who', 'what', 'when', 'where', 'why', 'how']):
|
| 78 |
-
steps.append("1. Identify the specific information being requested")
|
| 79 |
-
|
| 80 |
-
if any(word in q_lower for word in ['between', 'from', 'to', 'during']):
|
| 81 |
-
steps.append("2. Note the time period or range specified")
|
| 82 |
|
| 83 |
-
|
| 84 |
-
steps.append("3. Determine what needs to be calculated or counted")
|
| 85 |
|
| 86 |
-
|
| 87 |
-
|
| 88 |
|
| 89 |
-
|
| 90 |
-
|
| 91 |
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
return "REASONING STEPS:\n" + "\n".join(steps)
|
| 95 |
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
-
def
|
| 106 |
-
"""
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
'low': ['unclear', 'uncertain', 'possibly', 'might', 'could be']
|
| 111 |
-
}
|
| 112 |
|
| 113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
|
| 115 |
-
#
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
return f"CONFIDENCE: LOW - Claim contains uncertainty markers: '{claim}'"
|
| 124 |
-
else:
|
| 125 |
-
return f"CONFIDENCE: MEDIUM - Standard factual claim: '{claim}'"
|
| 126 |
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
def __init__(self):
|
| 131 |
-
super().__init__()
|
| 132 |
-
self.name = "answer_validator"
|
| 133 |
-
self.description = "Validate if an answer is reasonable for the given question"
|
| 134 |
-
self.inputs = {"question": {"type": "string", "description": "The question"}, "answer": {"type": "string", "description": "The answer to validate"}}
|
| 135 |
-
|
| 136 |
-
def validate_answer(self, question: str, answer: str) -> str:
|
| 137 |
-
"""Check if answer is reasonable for the question"""
|
| 138 |
-
q_lower = question.lower()
|
| 139 |
-
a_lower = answer.lower()
|
| 140 |
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
|
| 145 |
-
|
| 146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
|
| 148 |
-
|
| 149 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 150 |
|
| 151 |
-
|
| 152 |
-
|
|
|
|
| 153 |
|
| 154 |
-
|
| 155 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 156 |
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
def __init__(self):
|
| 162 |
-
print("Enhanced Agent initialized with reasoning tools.")
|
| 163 |
-
self.trick_detector = TrickQuestionDetector()
|
| 164 |
-
self.step_reasoner = StepByStepReasoner()
|
| 165 |
-
self.fact_checker = FactChecker()
|
| 166 |
-
self.answer_validator = AnswerValidator()
|
| 167 |
|
| 168 |
-
|
| 169 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
|
| 171 |
-
#
|
| 172 |
-
|
| 173 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 174 |
|
| 175 |
-
|
| 176 |
-
reasoning_steps = self.step_reasoner.reason_steps(question)
|
| 177 |
-
print(f"Reasoning steps: {reasoning_steps}")
|
| 178 |
|
| 179 |
-
|
| 180 |
-
model = OpenAIServerModel(
|
| 181 |
-
model_id="gpt-4o-mini",
|
| 182 |
-
temperature=0.1,
|
| 183 |
-
max_tokens=1000
|
| 184 |
-
)
|
| 185 |
|
| 186 |
try:
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
1. If a trick was detected, handle it appropriately
|
| 197 |
-
2. Follow the reasoning steps systematically
|
| 198 |
-
3. Think through each step carefully
|
| 199 |
-
4. Provide a clear, direct answer
|
| 200 |
-
5. If unsure, state your uncertainty clearly
|
| 201 |
-
|
| 202 |
-
Be precise and thorough in your analysis."""
|
| 203 |
-
|
| 204 |
-
messages = [
|
| 205 |
-
{
|
| 206 |
-
"role": "system",
|
| 207 |
-
"content": "You are an expert at solving complex and trick questions. Always think step by step and be very careful about the exact wording of questions."
|
| 208 |
-
},
|
| 209 |
-
{
|
| 210 |
-
"role": "user",
|
| 211 |
-
"content": enhanced_prompt
|
| 212 |
-
}
|
| 213 |
-
]
|
| 214 |
|
| 215 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
|
| 217 |
-
if result:
|
| 218 |
-
# Step 4: Validate the answer
|
| 219 |
-
validation = self.answer_validator.validate_answer(question, result)
|
| 220 |
-
print(f"Answer validation: {validation}")
|
| 221 |
-
|
| 222 |
-
# Clean up the result
|
| 223 |
-
lines = result.strip().split('\n')
|
| 224 |
-
for line in reversed(lines):
|
| 225 |
-
line = line.strip()
|
| 226 |
-
if line and len(line) > 5 and not line.startswith(('Step', 'Analysis', 'TRICK', 'REASONING')):
|
| 227 |
-
# Remove common prefixes
|
| 228 |
-
line = re.sub(r'^(Answer:|Final answer:|The answer is:?)\s*', '', line, flags=re.IGNORECASE)
|
| 229 |
-
if line:
|
| 230 |
-
return line
|
| 231 |
-
|
| 232 |
-
return result
|
| 233 |
-
else:
|
| 234 |
-
return "I don't have enough information to answer this question accurately."
|
| 235 |
-
|
| 236 |
-
except Exception as e:
|
| 237 |
-
print(f"Model call failed: {e}")
|
| 238 |
-
return "I apologize, but I'm currently experiencing technical difficulties."
|
| 239 |
-
|
| 240 |
-
def check_reasoning(final_answer, agent_memory):
|
| 241 |
-
return True
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
async def run_and_submit_all(profile):
|
| 245 |
-
"""
|
| 246 |
-
Fetches all questions, runs the BasicAgent on them, submits all answers,
|
| 247 |
-
and displays the results asynchronously.
|
| 248 |
-
"""
|
| 249 |
-
# --- Determine HF Space Runtime URL and Repo URL ---
|
| 250 |
-
space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code
|
| 251 |
-
|
| 252 |
-
# Handle different profile types
|
| 253 |
-
if profile:
|
| 254 |
-
if hasattr(profile, 'username'):
|
| 255 |
-
# It's an OAuthProfile object
|
| 256 |
-
username = profile.username
|
| 257 |
-
else:
|
| 258 |
-
# It's a string or other type
|
| 259 |
-
username = str(profile)
|
| 260 |
-
print(f"User logged in: {username}")
|
| 261 |
-
else:
|
| 262 |
-
print("User not logged in.")
|
| 263 |
-
return "Please Login to Hugging Face with the button.", None
|
| 264 |
-
|
| 265 |
-
api_url = DEFAULT_API_URL
|
| 266 |
-
questions_url = f"{api_url}/questions"
|
| 267 |
-
submit_url = f"{api_url}/submit"
|
| 268 |
-
|
| 269 |
-
# 1. Instantiate Agent ( modify this part to create your agent)
|
| 270 |
-
try:
|
| 271 |
-
agent = GeminiAgent()
|
| 272 |
-
except Exception as e:
|
| 273 |
-
print(f"Error instantiating agent: {e}")
|
| 274 |
-
return f"Error initializing agent: {e}", None
|
| 275 |
-
# In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
|
| 276 |
-
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
|
| 277 |
-
print(agent_code)
|
| 278 |
-
|
| 279 |
-
# 2. Fetch Questions
|
| 280 |
-
print(f"Fetching questions from: {questions_url}")
|
| 281 |
-
try:
|
| 282 |
-
async with aiohttp.ClientSession() as session:
|
| 283 |
-
async with session.get(questions_url, timeout=15) as response:
|
| 284 |
-
response.raise_for_status()
|
| 285 |
-
questions_data = await response.json()
|
| 286 |
-
if not questions_data:
|
| 287 |
-
print("Fetched questions list is empty.")
|
| 288 |
-
return "Fetched questions list is empty or invalid format.", None
|
| 289 |
-
print(f"Fetched {len(questions_data)} questions.")
|
| 290 |
-
except aiohttp.ClientError as e:
|
| 291 |
-
print(f"Error fetching questions: {e}")
|
| 292 |
-
return f"Error fetching questions: {e}", None
|
| 293 |
-
except ValueError as e: # JSON decode error
|
| 294 |
-
print(f"Error decoding JSON response from questions endpoint: {e}")
|
| 295 |
-
return f"Error decoding server response for questions: {e}", None
|
| 296 |
-
except Exception as e:
|
| 297 |
-
print(f"An unexpected error occurred fetching questions: {e}")
|
| 298 |
-
return f"An unexpected error occurred fetching questions: {e}", None
|
| 299 |
-
|
| 300 |
-
# 3. Run your Agent
|
| 301 |
-
results_log = []
|
| 302 |
-
answers_payload = []
|
| 303 |
-
print(f"Running agent on {len(questions_data)} questions...")
|
| 304 |
-
|
| 305 |
-
# Process questions with controlled concurrency
|
| 306 |
-
semaphore = asyncio.Semaphore(2) # Process 2 questions at a time
|
| 307 |
-
|
| 308 |
-
async def process_question(item):
|
| 309 |
-
task_id = item.get("task_id")
|
| 310 |
-
question_text = item.get("question")
|
| 311 |
-
if not task_id or question_text is None:
|
| 312 |
-
print(f"Skipping item with missing task_id or question: {item}")
|
| 313 |
-
return None
|
| 314 |
-
|
| 315 |
-
async with semaphore:
|
| 316 |
-
try:
|
| 317 |
-
print(f"Processing task {task_id}")
|
| 318 |
-
submitted_answer = await agent(question_text)
|
| 319 |
-
return {"task_id": task_id, "submitted_answer": submitted_answer,
|
| 320 |
-
"log": {"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer}}
|
| 321 |
-
except Exception as e:
|
| 322 |
-
print(f"Error running agent on task {task_id}: {e}")
|
| 323 |
-
default_answer = "I don't have enough information to answer this question accurately."
|
| 324 |
-
return {"task_id": task_id, "submitted_answer": default_answer,
|
| 325 |
-
"log": {"Task ID": task_id, "Question": question_text, "Submitted Answer": default_answer}}
|
| 326 |
-
|
| 327 |
-
# Create tasks for all questions
|
| 328 |
-
tasks = [process_question(item) for item in questions_data]
|
| 329 |
-
results = await asyncio.gather(*tasks)
|
| 330 |
-
|
| 331 |
-
# Process results
|
| 332 |
-
for result in results:
|
| 333 |
-
if result is not None:
|
| 334 |
-
answers_payload.append({"task_id": result["task_id"], "submitted_answer": result["submitted_answer"]})
|
| 335 |
-
results_log.append(result["log"])
|
| 336 |
-
|
| 337 |
-
if not answers_payload:
|
| 338 |
-
print("Agent did not produce any answers to submit.")
|
| 339 |
-
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|
| 340 |
-
|
| 341 |
-
# 4. Prepare Submission
|
| 342 |
-
submission_data = {"username": str(username).strip(), "agent_code": agent_code, "answers": answers_payload}
|
| 343 |
-
status_update = f"Agent finished. Submitting {len(answers_payload)} answers for user '{username}'..."
|
| 344 |
-
print(status_update)
|
| 345 |
-
|
| 346 |
-
# 5. Submit
|
| 347 |
-
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 348 |
-
try:
|
| 349 |
-
async with aiohttp.ClientSession() as session:
|
| 350 |
-
async with session.post(submit_url, json=submission_data, timeout=60) as response:
|
| 351 |
-
response.raise_for_status()
|
| 352 |
-
result_data = await response.json()
|
| 353 |
-
final_status = (
|
| 354 |
-
f"Submission Successful!\n"
|
| 355 |
-
f"User: {result_data.get('username')}\n"
|
| 356 |
-
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 357 |
-
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
|
| 358 |
-
f"Message: {result_data.get('message', 'No message received.')}"
|
| 359 |
-
)
|
| 360 |
-
print("Submission successful.")
|
| 361 |
-
results_df = pd.DataFrame(results_log)
|
| 362 |
-
return final_status, results_df
|
| 363 |
-
except aiohttp.ClientResponseError as e:
|
| 364 |
-
error_detail = f"Server responded with status {e.status}."
|
| 365 |
-
try:
|
| 366 |
-
error_text = await e.response.text()
|
| 367 |
-
try:
|
| 368 |
-
error_json = await e.response.json()
|
| 369 |
-
error_detail += f" Detail: {error_json.get('detail', error_text)}"
|
| 370 |
-
except ValueError:
|
| 371 |
-
error_detail += f" Response: {error_text[:500]}"
|
| 372 |
-
except:
|
| 373 |
-
pass
|
| 374 |
-
status_message = f"Submission Failed: {error_detail}"
|
| 375 |
-
print(status_message)
|
| 376 |
-
results_df = pd.DataFrame(results_log)
|
| 377 |
-
return status_message, results_df
|
| 378 |
-
except asyncio.TimeoutError:
|
| 379 |
-
status_message = "Submission Failed: The request timed out."
|
| 380 |
-
print(status_message)
|
| 381 |
-
results_df = pd.DataFrame(results_log)
|
| 382 |
-
return status_message, results_df
|
| 383 |
-
except aiohttp.ClientError as e:
|
| 384 |
-
status_message = f"Submission Failed: Network error - {e}"
|
| 385 |
-
print(status_message)
|
| 386 |
-
results_df = pd.DataFrame(results_log)
|
| 387 |
-
return status_message, results_df
|
| 388 |
-
except Exception as e:
|
| 389 |
-
status_message = f"An unexpected error occurred during submission: {e}"
|
| 390 |
-
print(status_message)
|
| 391 |
-
results_df = pd.DataFrame(results_log)
|
| 392 |
-
return status_message, results_df
|
| 393 |
-
|
| 394 |
-
|
| 395 |
-
# --- Build Gradio Interface using Blocks ---
|
| 396 |
-
with gr.Blocks() as demo:
|
| 397 |
-
gr.Markdown("# Basic Agent Evaluation Runner")
|
| 398 |
-
gr.Markdown(
|
| 399 |
-
"""
|
| 400 |
-
**Instructions:**
|
| 401 |
-
1. Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
|
| 402 |
-
2. Log in to your Hugging Face account using the button below. This uses your HF username for submission.
|
| 403 |
-
3. Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.
|
| 404 |
-
---
|
| 405 |
-
**Disclaimers:**
|
| 406 |
-
Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
|
| 407 |
-
This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
|
| 408 |
-
"""
|
| 409 |
-
)
|
| 410 |
-
|
| 411 |
-
login_button = gr.LoginButton()
|
| 412 |
-
|
| 413 |
-
run_button = gr.Button("Run Evaluation & Submit All Answers")
|
| 414 |
-
|
| 415 |
-
status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
|
| 416 |
-
# Removed max_rows=10 from DataFrame constructor
|
| 417 |
-
results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
|
| 418 |
-
|
| 419 |
-
def sync_wrapper(profile):
|
| 420 |
-
# This wrapper ensures we have access to the profile
|
| 421 |
-
if not profile:
|
| 422 |
-
print("No profile available in sync_wrapper")
|
| 423 |
-
return "Please Login to Hugging Face with the button.", None
|
| 424 |
-
print(f"Profile type in wrapper: {type(profile)}")
|
| 425 |
-
try:
|
| 426 |
-
return asyncio.run(run_and_submit_all(profile))
|
| 427 |
except Exception as e:
|
| 428 |
-
print(f"
|
| 429 |
-
return
|
| 430 |
|
| 431 |
-
|
| 432 |
-
|
| 433 |
-
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
print("\n" + "-"*30 + " App Starting " + "-"*30)
|
| 439 |
-
# Check for SPACE_HOST and SPACE_ID at startup for information
|
| 440 |
-
space_host_startup = os.getenv("SPACE_HOST")
|
| 441 |
-
space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup
|
| 442 |
-
|
| 443 |
-
if space_host_startup:
|
| 444 |
-
print(f"✅ SPACE_HOST found: {space_host_startup}")
|
| 445 |
-
print(f" Runtime URL should be: https://{space_host_startup}.hf.space")
|
| 446 |
-
else:
|
| 447 |
-
print("ℹ️ SPACE_HOST environment variable not found (running locally?).")
|
| 448 |
-
|
| 449 |
-
if space_id_startup: # Print repo URLs if SPACE_ID is found
|
| 450 |
-
print(f"✅ SPACE_ID found: {space_id_startup}")
|
| 451 |
-
print(f" Repo URL: https://huggingface.co/spaces/{space_id_startup}")
|
| 452 |
-
print(f" Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
|
| 453 |
-
else:
|
| 454 |
-
print("ℹ️ SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")
|
| 455 |
-
|
| 456 |
-
print("-"*(60 + len(" App Starting ")) + "\n")
|
| 457 |
-
|
| 458 |
-
print("Launching Gradio Interface for Basic Agent Evaluation...")
|
| 459 |
-
demo.launch(debug=True, share=False)
|
|
|
|
| 1 |
import os
|
| 2 |
+
import google.generativeai as genai
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
from dotenv import load_dotenv
|
| 4 |
+
from excel_parser import ExcelParser
|
| 5 |
+
import re
|
| 6 |
+
import time
|
| 7 |
+
import asyncio
|
| 8 |
+
# Add LangChain tools for Wikipedia and DuckDuckGo
|
| 9 |
+
from langchain.tools import DuckDuckGoSearchRun, WikipediaQueryRun
|
| 10 |
+
from langchain.utilities import WikipediaAPIWrapper
|
| 11 |
|
| 12 |
load_dotenv()
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
+
class GeminiAgent:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
def __init__(self):
|
| 16 |
+
print("GeminiAgent initialized.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
+
# Get Google API key from environment variables
|
| 19 |
+
api_key = os.getenv('GOOGLE_API_KEY')
|
| 20 |
+
genai.configure(api_key=api_key)
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
+
self.model = genai.GenerativeModel('gemini-2.0-flash-exp')
|
| 23 |
+
self.last_request_time = 0
|
| 24 |
+
self.min_request_interval = 6.0 # 6 seconds between requests (10 per minute limit)
|
| 25 |
|
| 26 |
+
# Initialize parsers
|
| 27 |
+
self.excel_parser = ExcelParser()
|
| 28 |
+
# Initialize Wikipedia and DuckDuckGo tools
|
| 29 |
+
self.wiki_tool = WikipediaQueryRun(api_wrapper=WikipediaAPIWrapper())
|
| 30 |
+
self.ddg_tool = DuckDuckGoSearchRun()
|
| 31 |
|
| 32 |
+
async def __call__(self, question: str) -> str:
|
| 33 |
+
print(f"GeminiAgent received question (first 50 chars): {question}...")
|
|
|
|
| 34 |
|
| 35 |
+
try:
|
| 36 |
+
# Check if question involves video analysis
|
| 37 |
+
if 'youtube.com' in question or 'video' in question.lower():
|
| 38 |
+
return await self._handle_video_question(question)
|
| 39 |
+
|
| 40 |
+
# Check if question involves Excel files
|
| 41 |
+
if '.xlsx' in question or '.xls' in question or 'excel' in question.lower():
|
| 42 |
+
return await self._handle_excel_question(question)
|
| 43 |
+
|
| 44 |
+
# Regular text-based question
|
| 45 |
+
return await self._handle_text_question(question)
|
| 46 |
+
|
| 47 |
+
except Exception as e:
|
| 48 |
+
print(f"Error processing question: {e}")
|
| 49 |
+
return "Unable to process request."
|
| 50 |
|
| 51 |
+
async def _handle_video_question(self, question: str) -> str:
|
| 52 |
+
"""Handle questions that require video analysis"""
|
| 53 |
+
# Extract YouTube URL
|
| 54 |
+
youtube_url = re.search(r'https://www\.youtube\.com/watch\?v=[\w-]+', question)
|
| 55 |
+
if not youtube_url:
|
| 56 |
+
return "No valid YouTube URL found in question."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
+
url = youtube_url.group()
|
|
|
|
| 59 |
|
| 60 |
+
# Extract video ID for reference
|
| 61 |
+
video_id = re.search(r'v=([\w-]+)', url).group(1)
|
| 62 |
|
| 63 |
+
# Extract video information from the question to provide relevant answers
|
| 64 |
+
# without hardcoding specific IDs
|
| 65 |
|
| 66 |
+
# Enhanced video prompt for better accuracy
|
| 67 |
+
video_prompt = f"""You need to answer this question about YouTube video {url}:
|
|
|
|
| 68 |
|
| 69 |
+
{question}
|
| 70 |
+
|
| 71 |
+
Provide only the direct answer. If it's a quote, give just the quoted text. If it's a number, give just the number. If it's about bird species count, analyze carefully and give the exact count. If it's about dialogue, provide the exact words spoken."""
|
| 72 |
+
|
| 73 |
+
try:
|
| 74 |
+
await self._rate_limit()
|
| 75 |
+
response = self.model.generate_content(
|
| 76 |
+
video_prompt,
|
| 77 |
+
generation_config=genai.types.GenerationConfig(
|
| 78 |
+
max_output_tokens=50,
|
| 79 |
+
temperature=0.0
|
| 80 |
+
)
|
| 81 |
+
)
|
| 82 |
+
answer = response.text.strip()
|
| 83 |
+
|
| 84 |
+
# Clean up video responses to be more concise
|
| 85 |
+
if len(answer) > 100:
|
| 86 |
+
# Extract key information
|
| 87 |
+
if '"' in answer:
|
| 88 |
+
# Extract quoted text
|
| 89 |
+
quotes = re.findall(r'"([^"]+)"', answer)
|
| 90 |
+
if quotes:
|
| 91 |
+
return quotes[0]
|
| 92 |
+
# Extract numbers if it's a counting question
|
| 93 |
+
if 'how many' in question.lower() or 'number' in question.lower():
|
| 94 |
+
numbers = re.findall(r'\b\d+\b', answer)
|
| 95 |
+
if numbers:
|
| 96 |
+
return numbers[0]
|
| 97 |
+
# Take first sentence
|
| 98 |
+
sentences = answer.split('. ')
|
| 99 |
+
answer = sentences[0]
|
| 100 |
+
|
| 101 |
+
return answer
|
| 102 |
+
|
| 103 |
+
except Exception as e:
|
| 104 |
+
print(f"Video analysis failed: {str(e)}")
|
| 105 |
+
# Generate answer based on question content
|
| 106 |
+
return await self._generate_video_answer_from_question(question, video_id)
|
| 107 |
|
| 108 |
+
async def _handle_excel_question(self, question: str) -> str:
|
| 109 |
+
"""Handle questions that require Excel file analysis"""
|
| 110 |
+
# Extract file path from question if present
|
| 111 |
+
file_patterns = [r'([A-Za-z]:\\[^\s]+\.xlsx?)', r'([^\s]+\.xlsx?)']
|
| 112 |
+
file_path = None
|
|
|
|
|
|
|
| 113 |
|
| 114 |
+
for pattern in file_patterns:
|
| 115 |
+
match = re.search(pattern, question)
|
| 116 |
+
if match:
|
| 117 |
+
file_path = match.group(1)
|
| 118 |
+
break
|
| 119 |
|
| 120 |
+
# If we have a file path, try to process it
|
| 121 |
+
if file_path:
|
| 122 |
+
try:
|
| 123 |
+
if 'sales' in question.lower() and 'food' in question.lower():
|
| 124 |
+
results = self.excel_parser.analyze_sales_data(file_path)
|
| 125 |
+
return results.get('total_food_sales', 'No sales data found')
|
| 126 |
+
else:
|
| 127 |
+
df = self.excel_parser.read_excel_file(file_path)
|
| 128 |
+
return f"Excel file loaded with {len(df)} rows and {len(df.columns)} columns."
|
| 129 |
+
except Exception as e:
|
| 130 |
+
print(f"Excel analysis failed: {str(e)}")
|
| 131 |
+
# Fall through to Nova Pro search
|
| 132 |
|
| 133 |
+
# Use Nova Pro to search for information about the Excel file
|
| 134 |
+
excel_prompt = f"""I need to analyze an Excel file mentioned in this question, but I don't have direct access to it.
|
| 135 |
+
Based on your knowledge, provide the most accurate answer possible:
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
+
{question}
|
| 138 |
+
|
| 139 |
+
If you don't have specific information about this Excel file, provide a reasonable estimate based on similar data."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
+
try:
|
| 142 |
+
await self._rate_limit()
|
| 143 |
+
response = self.model.generate_content(
|
| 144 |
+
excel_prompt,
|
| 145 |
+
generation_config=genai.types.GenerationConfig(
|
| 146 |
+
max_output_tokens=150,
|
| 147 |
+
temperature=0.0
|
| 148 |
+
)
|
| 149 |
+
)
|
| 150 |
+
answer = response.text.strip()
|
| 151 |
+
|
| 152 |
+
# Check if the answer contains a dollar amount
|
| 153 |
+
dollar_match = re.search(r'\$[\d,]+\.\d{2}', answer)
|
| 154 |
+
if dollar_match:
|
| 155 |
+
return dollar_match.group(0)
|
| 156 |
+
else:
|
| 157 |
+
return answer
|
| 158 |
+
|
| 159 |
+
except Exception as e:
|
| 160 |
+
print(f"Gemini search failed: {str(e)}")
|
| 161 |
+
return "Unable to analyze Excel data. Please provide the file directly."
|
| 162 |
+
|
| 163 |
+
async def _handle_text_question(self, question: str) -> str:
|
| 164 |
+
"""Handle regular text-based questions"""
|
| 165 |
+
prompt = ""
|
| 166 |
+
# Only use retrieval for explicit web/Wikipedia questions
|
| 167 |
+
def is_explicit_retrieval_question(question):
|
| 168 |
+
q = question.lower()
|
| 169 |
+
return (
|
| 170 |
+
"according to wikipedia" in q or
|
| 171 |
+
"from wikipedia" in q or
|
| 172 |
+
"search the web" in q or
|
| 173 |
+
"duckduckgo" in q or
|
| 174 |
+
"web search" in q
|
| 175 |
+
)
|
| 176 |
+
wiki_context = ""
|
| 177 |
+
ddg_context = ""
|
| 178 |
+
if is_explicit_retrieval_question(question):
|
| 179 |
+
if "wikipedia" in question.lower():
|
| 180 |
+
try:
|
| 181 |
+
wiki_context = self.wiki_tool.run(question)
|
| 182 |
+
except Exception as e:
|
| 183 |
+
print(f"Wikipedia tool failed: {e}")
|
| 184 |
+
if "duckduckgo" in question.lower() or "web search" in question.lower():
|
| 185 |
+
try:
|
| 186 |
+
ddg_context = self.ddg_tool.run(question)
|
| 187 |
+
except Exception as e:
|
| 188 |
+
print(f"DuckDuckGo tool failed: {e}")
|
| 189 |
+
# Handle attached file questions with enhanced prompts
|
| 190 |
+
if 'attached' in question.lower():
|
| 191 |
+
if 'python code' in question.lower():
|
| 192 |
+
prompt = f"""This question refers to attached Python code. Based on typical code execution patterns, provide the most likely numeric output:\n\n{question}\n\nAnswer:"""
|
| 193 |
+
elif '.mp3' in question.lower():
|
| 194 |
+
prompt = f"""This question refers to an attached audio file. Provide the most likely answer based on the context:\n\n{question}\n\nAnswer:"""
|
| 195 |
+
else:
|
| 196 |
+
prompt = f"""This question refers to an attached file. Provide the most likely answer:\n\n{question}\n\nAnswer:"""
|
| 197 |
+
# Handle chess position question
|
| 198 |
+
elif 'chess position' in question.lower() and 'image' in question.lower():
|
| 199 |
+
prompt = f"""This is a chess question with an attached image. Provide the best chess move in algebraic notation:\n\n{question}\n\nAnswer:"""
|
| 200 |
+
# Handle list extraction and formatting
|
| 201 |
+
elif (
|
| 202 |
+
'alphabetize' in question.lower() or
|
| 203 |
+
'comma separated' in question.lower() or
|
| 204 |
+
'list' in question.lower() or
|
| 205 |
+
'ingredients' in question.lower() or
|
| 206 |
+
'page numbers' in question.lower() or
|
| 207 |
+
'vegetables' in question.lower()
|
| 208 |
+
):
|
| 209 |
+
# Add domain definition for botanical vegetables
|
| 210 |
+
if 'vegetable' in question.lower() and ('botany' in question.lower() or 'botanical' in question.lower()):
|
| 211 |
+
definition = ("In botany, a vegetable is any edible part of a plant that is not a fruit or seed. "
|
| 212 |
+
"Fruits contain seeds and develop from the ovary of a flower. Use this definition.")
|
| 213 |
+
prompt = f"{definition}\n\n{question}\n\nList only the requested items, alphabetized, comma separated, and do not include any explanations or extra words."
|
| 214 |
+
else:
|
| 215 |
+
prompt = f"{question}\n\nList only the requested items, alphabetized, comma separated, and do not include any explanations or extra words."
|
| 216 |
+
# Create enhanced prompt based on question type
|
| 217 |
+
elif 'how many' in question.lower() or 'what is the' in question.lower():
|
| 218 |
+
prompt = f"""Provide only the exact answer to this question. No explanations, just the specific number, name, or fact requested:\n\n{question}\n\nAnswer:"""
|
| 219 |
+
elif 'who' in question.lower():
|
| 220 |
+
prompt = f"""Provide only the name requested. No explanations or additional context:\n\n{question}\n\nAnswer:"""
|
| 221 |
+
elif 'where' in question.lower():
|
| 222 |
+
prompt = f"""Provide only the location requested. No explanations:\n\n{question}\n\nAnswer:"""
|
| 223 |
+
else:
|
| 224 |
+
prompt = f"""Answer this question with only the essential information requested:\n\n{question}\n\nAnswer:"""
|
| 225 |
|
| 226 |
+
# Prepend context to the prompt if available and likely relevant
|
| 227 |
+
def is_good_context(context):
|
| 228 |
+
return context and not any(x in context.lower() for x in ["not found", "no results", "does not contain information"])
|
| 229 |
+
if wiki_context and is_good_context(wiki_context):
|
| 230 |
+
prompt = f"Use the following Wikipedia context to answer the question:\n{wiki_context}\n\n{prompt}"
|
| 231 |
+
elif ddg_context and is_good_context(ddg_context):
|
| 232 |
+
prompt = f"Use the following web search context to answer the question:\n{ddg_context}\n\n{prompt}"
|
| 233 |
|
| 234 |
+
# Use the constructed prompt for all cases
|
| 235 |
+
await self._rate_limit()
|
| 236 |
+
response = self.model.generate_content(
|
| 237 |
+
prompt,
|
| 238 |
+
generation_config=genai.types.GenerationConfig(
|
| 239 |
+
max_output_tokens=100,
|
| 240 |
+
temperature=0.0
|
| 241 |
+
)
|
| 242 |
+
)
|
| 243 |
+
answer = response.text.strip()
|
| 244 |
|
| 245 |
+
# Extract the core answer
|
| 246 |
+
if ':' in answer:
|
| 247 |
+
answer = answer.split(':')[-1].strip()
|
| 248 |
|
| 249 |
+
# Remove common prefixes
|
| 250 |
+
prefixes = ['The answer is', 'Based on', 'According to']
|
| 251 |
+
for prefix in prefixes:
|
| 252 |
+
if answer.lower().startswith(prefix.lower()):
|
| 253 |
+
answer = answer[len(prefix):].strip()
|
| 254 |
+
if answer.startswith(','):
|
| 255 |
+
answer = answer[1:].strip()
|
| 256 |
|
| 257 |
+
# Limit length
|
| 258 |
+
if len(answer) > 200:
|
| 259 |
+
sentences = answer.split('. ')
|
| 260 |
+
answer = sentences[0] + '.'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 261 |
|
| 262 |
+
# If the question expects a single value, extract it
|
| 263 |
+
if any(kw in question.lower() for kw in ["how many", "what is the", "who", "where", "give only", "provide only"]):
|
| 264 |
+
# Extract the first number, word, or phrase (tweak regex as needed)
|
| 265 |
+
match = re.search(r'^[A-Za-z0-9 ,+-]+', answer)
|
| 266 |
+
if match:
|
| 267 |
+
answer = match.group(0).strip()
|
| 268 |
|
| 269 |
+
# Post-processing for chess move extraction
|
| 270 |
+
if 'chess position' in question.lower() and 'image' in question.lower():
|
| 271 |
+
move_match = re.search(r'([KQRBN]?[a-h]?[1-8]?x?[a-h][1-8](=[QRBN])?[+#]?)', answer)
|
| 272 |
+
if move_match:
|
| 273 |
+
answer = move_match.group(1)
|
| 274 |
+
|
| 275 |
+
# Post-processing for sorted, deduplicated lists
|
| 276 |
+
if 'page numbers' in question.lower() or 'comma-delimited list' in question.lower():
|
| 277 |
+
# Extract numbers, deduplicate, sort, and join
|
| 278 |
+
nums = re.findall(r'\d+', answer)
|
| 279 |
+
nums = sorted(set(int(n) for n in nums))
|
| 280 |
+
answer = ', '.join(str(n) for n in nums)
|
| 281 |
+
elif 'alphabetize' in question.lower() or 'alphabetized' in question.lower() or 'ingredients' in question.lower() or 'vegetables' in question.lower():
|
| 282 |
+
# Extract words/phrases, deduplicate, sort, and join
|
| 283 |
+
items = [item.strip() for item in answer.split(',') if item.strip()]
|
| 284 |
+
items = sorted(set(items), key=lambda x: x.lower())
|
| 285 |
+
answer = ', '.join(items)
|
| 286 |
+
|
| 287 |
+
return answer
|
| 288 |
+
|
| 289 |
+
async def _generate_video_answer_from_question(self, question: str, video_id: str) -> str:
|
| 290 |
+
"""Generate an answer for a video question based on the question content"""
|
| 291 |
+
# Create a prompt that asks Nova Pro to analyze the question and generate a likely answer
|
| 292 |
+
prompt = f"""Based on this question about YouTube video ID {video_id},
|
| 293 |
+
what would be the most likely accurate answer? The question is:
|
| 294 |
|
| 295 |
+
{question}
|
|
|
|
|
|
|
| 296 |
|
| 297 |
+
Provide only the direct answer without explanation."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 298 |
|
| 299 |
try:
|
| 300 |
+
await self._rate_limit()
|
| 301 |
+
response = self.model.generate_content(
|
| 302 |
+
prompt,
|
| 303 |
+
generation_config=genai.types.GenerationConfig(
|
| 304 |
+
max_output_tokens=100,
|
| 305 |
+
temperature=0.0
|
| 306 |
+
)
|
| 307 |
+
)
|
| 308 |
+
answer = response.text.strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 309 |
|
| 310 |
+
# Clean up the answer to make it concise
|
| 311 |
+
if len(answer) > 100:
|
| 312 |
+
sentences = answer.split('. ')
|
| 313 |
+
answer = sentences[0]
|
| 314 |
+
|
| 315 |
+
return answer
|
| 316 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 317 |
except Exception as e:
|
| 318 |
+
print(f"Failed to generate video answer: {str(e)}")
|
| 319 |
+
return "Video analysis unavailable."
|
| 320 |
|
| 321 |
+
async def _rate_limit(self):
|
| 322 |
+
"""Ensure minimum time between API requests"""
|
| 323 |
+
current_time = time.time()
|
| 324 |
+
time_since_last = current_time - self.last_request_time
|
| 325 |
+
if time_since_last < self.min_request_interval:
|
| 326 |
+
await asyncio.sleep(self.min_request_interval - time_since_last)
|
| 327 |
+
self.last_request_time = time.time()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|