Spaces:
Runtime error
Runtime error
Omachoko
commited on
Commit
·
83a3deb
1
Parent(s):
a9d900f
Final: robust GAIA agent with advanced tool registry, GPT-4.1, web search, strict output, and full multi-modal support
Browse files
app.py
CHANGED
@@ -24,6 +24,7 @@ from huggingface_hub import InferenceClient
|
|
24 |
import cv2
|
25 |
import torch
|
26 |
from bs4 import BeautifulSoup
|
|
|
27 |
|
28 |
logging.basicConfig(filename='gaia_agent.log', level=logging.INFO, format='%(asctime)s %(levelname)s:%(message)s')
|
29 |
logger = logging.getLogger(__name__)
|
@@ -198,6 +199,37 @@ def youtube_video_qa(youtube_url, question):
|
|
198 |
logging.error(f"YouTube video QA error: {e}")
|
199 |
return f"Video analysis error: {e}"
|
200 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
201 |
TOOL_REGISTRY = {
|
202 |
"llama3_chat": llama3_chat,
|
203 |
"mixtral_chat": mixtral_chat,
|
@@ -207,6 +239,8 @@ TOOL_REGISTRY = {
|
|
207 |
"image_caption": image_caption,
|
208 |
"code_analysis": code_analysis,
|
209 |
"youtube_video_qa": youtube_video_qa,
|
|
|
|
|
210 |
}
|
211 |
|
212 |
class ModularGAIAAgent:
|
@@ -304,63 +338,80 @@ class ModularGAIAAgent:
|
|
304 |
self.reasoning_trace.append(f"Unknown file type: {file_name}")
|
305 |
return None
|
306 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
307 |
def answer_question(self, question_obj):
|
308 |
self.reasoning_trace = []
|
309 |
q = question_obj["question"]
|
310 |
file_name = question_obj.get("file_name", "")
|
311 |
file_content = None
|
312 |
file_type = None
|
313 |
-
# YouTube video question detection
|
314 |
-
if "youtube.com" in q or "youtu.be" in q:
|
315 |
-
url = None
|
316 |
-
for word in q.split():
|
317 |
-
if "youtube.com" in word or "youtu.be" in word:
|
318 |
-
url = word.strip().strip(',')
|
319 |
-
break
|
320 |
-
if url:
|
321 |
-
answer = self.tools['youtube_video_qa'](url, q)
|
322 |
-
self.reasoning_trace.append(f"YouTube video analyzed: {url}")
|
323 |
-
self.reasoning_trace.append(f"Final answer: {answer}")
|
324 |
-
return self.format_answer(answer), self.reasoning_trace
|
325 |
if file_name:
|
326 |
file_id = file_name.split('.')[0]
|
327 |
local_file = self.download_file(file_id, file_name)
|
328 |
if local_file:
|
329 |
file_type = self.detect_file_type(local_file)
|
330 |
file_content = self.analyze_file(local_file, file_type)
|
331 |
-
#
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
answer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
340 |
else:
|
341 |
-
answer =
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
else:
|
346 |
-
answer = self.tools['llama3_chat'](q)
|
347 |
-
elif file_type == 'code':
|
348 |
-
answer = file_content
|
349 |
-
else:
|
350 |
-
answer = self.tools['llama3_chat'](q)
|
351 |
self.reasoning_trace.append(f"Final answer: {answer}")
|
352 |
return self.format_answer(answer), self.reasoning_trace
|
353 |
|
354 |
def format_answer(self, answer):
|
|
|
355 |
if isinstance(answer, str):
|
356 |
-
|
357 |
-
|
358 |
-
if answer.lower().startswith(prefix):
|
359 |
-
answer = answer[len(prefix):].strip()
|
360 |
-
import re
|
361 |
-
answer = re.sub(r'\b(the|a|an)\b ', '', answer, flags=re.IGNORECASE)
|
362 |
-
answer = answer.strip().rstrip('.')
|
363 |
-
return answer
|
364 |
|
365 |
# --- Basic Agent Definition (now wraps ModularGAIAAgent) ---
|
366 |
class BasicAgent:
|
|
|
24 |
import cv2
|
25 |
import torch
|
26 |
from bs4 import BeautifulSoup
|
27 |
+
import openai
|
28 |
|
29 |
logging.basicConfig(filename='gaia_agent.log', level=logging.INFO, format='%(asctime)s %(levelname)s:%(message)s')
|
30 |
logger = logging.getLogger(__name__)
|
|
|
199 |
logging.error(f"YouTube video QA error: {e}")
|
200 |
return f"Video analysis error: {e}"
|
201 |
|
202 |
+
def web_search_duckduckgo(query, max_results=5):
|
203 |
+
"""DuckDuckGo web search tool: returns top snippets and URLs."""
|
204 |
+
try:
|
205 |
+
import duckduckgo_search
|
206 |
+
results = duckduckgo_search.DuckDuckGoSearch().search(query, max_results=max_results)
|
207 |
+
snippets = []
|
208 |
+
for r in results:
|
209 |
+
snippet = f"Title: {r['title']}\nSnippet: {r['body']}\nURL: {r['href']}"
|
210 |
+
snippets.append(snippet)
|
211 |
+
return '\n---\n'.join(snippets)
|
212 |
+
except Exception as e:
|
213 |
+
logging.error(f"web_search_duckduckgo error: {e}")
|
214 |
+
return f"Web search error: {e}"
|
215 |
+
|
216 |
+
def gpt4_chat(prompt, api_key=None):
|
217 |
+
"""OpenAI GPT-4.1 chat completion."""
|
218 |
+
try:
|
219 |
+
api_key = api_key or os.environ.get("OPENAI_API_KEY", "")
|
220 |
+
if not api_key:
|
221 |
+
return "No OpenAI API key provided."
|
222 |
+
response = openai.ChatCompletion.create(
|
223 |
+
model="gpt-4-1106-preview",
|
224 |
+
messages=[{"role": "system", "content": "You are a general AI assistant. Answer using as few words as possible, in the required format. Use tools as needed, and only output the answer."},
|
225 |
+
{"role": "user", "content": prompt}],
|
226 |
+
api_key=api_key,
|
227 |
+
)
|
228 |
+
return response.choices[0].message['content'].strip()
|
229 |
+
except Exception as e:
|
230 |
+
logging.error(f"gpt4_chat error: {e}")
|
231 |
+
return f"GPT-4 error: {e}"
|
232 |
+
|
233 |
TOOL_REGISTRY = {
|
234 |
"llama3_chat": llama3_chat,
|
235 |
"mixtral_chat": mixtral_chat,
|
|
|
239 |
"image_caption": image_caption,
|
240 |
"code_analysis": code_analysis,
|
241 |
"youtube_video_qa": youtube_video_qa,
|
242 |
+
"web_search_duckduckgo": web_search_duckduckgo,
|
243 |
+
"gpt4_chat": gpt4_chat,
|
244 |
}
|
245 |
|
246 |
class ModularGAIAAgent:
|
|
|
338 |
self.reasoning_trace.append(f"Unknown file type: {file_name}")
|
339 |
return None
|
340 |
|
341 |
+
def smart_tool_select(self, question, file_type=None):
|
342 |
+
"""Select the best tool(s) for the question, optionally using GPT-4.1 for planning."""
|
343 |
+
# Use GPT-4.1 to suggest a tool if available
|
344 |
+
api_key = os.environ.get("OPENAI_API_KEY", "")
|
345 |
+
if api_key:
|
346 |
+
plan_prompt = f"""
|
347 |
+
You are an expert AI agent. Given the following question and file type, suggest the best tool(s) to use from this list: {list(self.tools.keys())}.
|
348 |
+
Question: {question}
|
349 |
+
File type: {file_type}
|
350 |
+
Respond with a comma-separated list of tool names only, in order of use. If unsure, start with web_search_duckduckgo.
|
351 |
+
"""
|
352 |
+
plan = gpt4_chat(plan_prompt, api_key=api_key)
|
353 |
+
tool_names = [t.strip() for t in plan.split(',') if t.strip() in self.tools]
|
354 |
+
if tool_names:
|
355 |
+
return tool_names
|
356 |
+
# Fallback: heuristic
|
357 |
+
if file_type == 'audio':
|
358 |
+
return ['asr_transcribe']
|
359 |
+
elif file_type == 'image':
|
360 |
+
return ['image_caption']
|
361 |
+
elif file_type == 'code':
|
362 |
+
return ['code_analysis']
|
363 |
+
elif file_type in ['excel', 'csv']:
|
364 |
+
return ['table_qa']
|
365 |
+
elif 'youtube.com' in question or 'youtu.be' in question:
|
366 |
+
return ['youtube_video_qa']
|
367 |
+
elif any(w in question.lower() for w in ['wikipedia', 'who', 'when', 'where', 'what', 'how', 'find', 'search']):
|
368 |
+
return ['web_search_duckduckgo']
|
369 |
+
else:
|
370 |
+
return ['llama3_chat']
|
371 |
+
|
372 |
def answer_question(self, question_obj):
|
373 |
self.reasoning_trace = []
|
374 |
q = question_obj["question"]
|
375 |
file_name = question_obj.get("file_name", "")
|
376 |
file_content = None
|
377 |
file_type = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
378 |
if file_name:
|
379 |
file_id = file_name.split('.')[0]
|
380 |
local_file = self.download_file(file_id, file_name)
|
381 |
if local_file:
|
382 |
file_type = self.detect_file_type(local_file)
|
383 |
file_content = self.analyze_file(local_file, file_type)
|
384 |
+
# Smart tool selection
|
385 |
+
tool_names = self.smart_tool_select(q, file_type)
|
386 |
+
answer = None
|
387 |
+
context = None
|
388 |
+
for tool_name in tool_names:
|
389 |
+
tool = self.tools[tool_name]
|
390 |
+
if tool_name == 'web_search_duckduckgo':
|
391 |
+
context = tool(q)
|
392 |
+
# Use LLM to synthesize answer from snippets
|
393 |
+
answer = llama3_chat(f"Answer the following question using ONLY the information below.\nQuestion: {q}\nSnippets:\n{context}\nAnswer:")
|
394 |
+
elif tool_name == 'gpt4_chat':
|
395 |
+
answer = tool(q)
|
396 |
+
elif tool_name == 'table_qa' and file_content:
|
397 |
+
answer = tool(q, file_content)
|
398 |
+
elif tool_name in ['asr_transcribe', 'image_caption', 'code_analysis'] and file_content:
|
399 |
+
answer = tool(file_name)
|
400 |
+
elif tool_name == 'youtube_video_qa':
|
401 |
+
answer = tool(q, q)
|
402 |
else:
|
403 |
+
answer = tool(q)
|
404 |
+
if answer:
|
405 |
+
break
|
406 |
+
self.reasoning_trace.append(f"Tools used: {tool_names}")
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
self.reasoning_trace.append(f"Final answer: {answer}")
|
408 |
return self.format_answer(answer), self.reasoning_trace
|
409 |
|
410 |
def format_answer(self, answer):
|
411 |
+
# Strict GAIA: only the answer, no extra text, no prefix
|
412 |
if isinstance(answer, str):
|
413 |
+
return answer.strip().split('\n')[0]
|
414 |
+
return str(answer)
|
|
|
|
|
|
|
|
|
|
|
|
|
415 |
|
416 |
# --- Basic Agent Definition (now wraps ModularGAIAAgent) ---
|
417 |
class BasicAgent:
|