Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -8,6 +8,7 @@ from zipfile import ZipFile
|
|
| 8 |
import gradio as gr
|
| 9 |
import pandas as pd
|
| 10 |
import requests
|
|
|
|
| 11 |
from smolagents import (
|
| 12 |
CodeAgent,
|
| 13 |
DuckDuckGoSearchTool,
|
|
@@ -22,9 +23,8 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
| 22 |
RESULT_COLUMNS = ["Task ID", "Question", "Submitted Answer"]
|
| 23 |
HTTP_TIMEOUT = 45
|
| 24 |
MAX_EXTRACTED_CHARS = 35_000
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
)
|
| 28 |
|
| 29 |
|
| 30 |
def clean_filename(value: str) -> str:
|
|
@@ -308,36 +308,107 @@ class AnalyzeGaiaImageTool(Tool):
|
|
| 308 |
return f"Could not analyze the GAIA image: {exc}"
|
| 309 |
|
| 310 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 311 |
class BasicAgent:
|
| 312 |
def __init__(self):
|
| 313 |
print("Inicializando o agente GAIA...")
|
| 314 |
|
| 315 |
hf_token = os.getenv("HF_TOKEN")
|
| 316 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 317 |
raise RuntimeError(
|
| 318 |
-
"O secret
|
| 319 |
-
"
|
| 320 |
-
"Settings > Secrets."
|
| 321 |
)
|
| 322 |
|
| 323 |
self.model = LiteLLMModel(
|
| 324 |
-
model_id=
|
| 325 |
-
api_key=
|
| 326 |
temperature=0,
|
|
|
|
| 327 |
)
|
| 328 |
self.hf_token = hf_token
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 329 |
self.agent = CodeAgent(
|
| 330 |
-
tools=
|
| 331 |
-
DuckDuckGoSearchTool(max_results=8, rate_limit=1.0),
|
| 332 |
-
VisitWebpageTool(max_output_length=30_000),
|
| 333 |
-
WikipediaSearchTool(
|
| 334 |
-
user_agent="GAIA-Course-Agent/1.0 (educational project)",
|
| 335 |
-
language="en",
|
| 336 |
-
),
|
| 337 |
-
InspectGaiaAttachmentTool(),
|
| 338 |
-
YouTubeTranscriptTool(),
|
| 339 |
-
AnalyzeGaiaImageTool(),
|
| 340 |
-
],
|
| 341 |
model=self.model,
|
| 342 |
max_steps=10,
|
| 343 |
planning_interval=4,
|
|
@@ -367,6 +438,10 @@ If the task asks what someone says in a YouTube video, call youtube_transcript
|
|
| 367 |
with the exact video URL before searching the web.
|
| 368 |
If the task depends on an attached image, call analyze_gaia_image with the
|
| 369 |
task_id and complete question. Do not try to infer image contents from metadata.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 370 |
Prefer primary or official sources. When search snippets conflict, open the
|
| 371 |
source and verify the relevant passage instead of guessing.
|
| 372 |
Only call tools that are explicitly available. Never invent a function such as
|
|
@@ -404,9 +479,9 @@ If a number is requested, return only that number.
|
|
| 404 |
or "Unauthorized" in error_text
|
| 405 |
):
|
| 406 |
raise RuntimeError(
|
| 407 |
-
"Falha de autenticação no modelo.
|
| 408 |
-
"
|
| 409 |
-
"
|
| 410 |
) from exc
|
| 411 |
raise
|
| 412 |
return self.format_exact_answer(question, str(result))
|
|
|
|
| 8 |
import gradio as gr
|
| 9 |
import pandas as pd
|
| 10 |
import requests
|
| 11 |
+
from litellm import completion
|
| 12 |
from smolagents import (
|
| 13 |
CodeAgent,
|
| 14 |
DuckDuckGoSearchTool,
|
|
|
|
| 23 |
RESULT_COLUMNS = ["Task ID", "Question", "Submitted Answer"]
|
| 24 |
HTTP_TIMEOUT = 45
|
| 25 |
MAX_EXTRACTED_CHARS = 35_000
|
| 26 |
+
DEFAULT_HF_MODEL = "huggingface/openai/gpt-oss-120b"
|
| 27 |
+
DEFAULT_GEMINI_MODEL = "gemini/gemini-3.5-flash"
|
|
|
|
| 28 |
|
| 29 |
|
| 30 |
def clean_filename(value: str) -> str:
|
|
|
|
| 308 |
return f"Could not analyze the GAIA image: {exc}"
|
| 309 |
|
| 310 |
|
| 311 |
+
class ConsultGeminiTool(Tool):
|
| 312 |
+
name = "consult_gemini"
|
| 313 |
+
description = (
|
| 314 |
+
"Asks Gemini for an independent second opinion on a difficult question "
|
| 315 |
+
"or on evidence already collected. Use it to verify reasoning, resolve "
|
| 316 |
+
"conflicting sources, or check the exact requested answer format. Do "
|
| 317 |
+
"not use it as a substitute for inspecting official attachments."
|
| 318 |
+
)
|
| 319 |
+
inputs = {
|
| 320 |
+
"request": {
|
| 321 |
+
"type": "string",
|
| 322 |
+
"description": (
|
| 323 |
+
"The complete question plus any relevant evidence and the "
|
| 324 |
+
"specific point Gemini should verify."
|
| 325 |
+
),
|
| 326 |
+
}
|
| 327 |
+
}
|
| 328 |
+
output_type = "string"
|
| 329 |
+
|
| 330 |
+
def forward(self, request: str) -> str:
|
| 331 |
+
api_key = os.getenv("GEMINI_API_KEY")
|
| 332 |
+
if not api_key:
|
| 333 |
+
return "Gemini consultation unavailable: GEMINI_API_KEY is missing."
|
| 334 |
+
|
| 335 |
+
model = os.getenv(
|
| 336 |
+
"GAIA_GEMINI_TOOL_MODEL", DEFAULT_GEMINI_MODEL
|
| 337 |
+
)
|
| 338 |
+
try:
|
| 339 |
+
response = completion(
|
| 340 |
+
model=model,
|
| 341 |
+
api_key=api_key,
|
| 342 |
+
messages=[
|
| 343 |
+
{
|
| 344 |
+
"role": "system",
|
| 345 |
+
"content": (
|
| 346 |
+
"You are a verification specialist assisting another "
|
| 347 |
+
"GAIA agent. Analyze the supplied question and evidence "
|
| 348 |
+
"critically. Identify uncertainty or contradictions, "
|
| 349 |
+
"then provide your recommended exact answer. Be concise "
|
| 350 |
+
"and never claim to have opened sources that were not "
|
| 351 |
+
"included in the request."
|
| 352 |
+
),
|
| 353 |
+
},
|
| 354 |
+
{"role": "user", "content": str(request)},
|
| 355 |
+
],
|
| 356 |
+
temperature=0,
|
| 357 |
+
max_tokens=900,
|
| 358 |
+
)
|
| 359 |
+
return str(response.choices[0].message.content).strip()
|
| 360 |
+
except Exception as exc:
|
| 361 |
+
return f"Gemini consultation failed: {exc}"
|
| 362 |
+
|
| 363 |
+
|
| 364 |
class BasicAgent:
|
| 365 |
def __init__(self):
|
| 366 |
print("Inicializando o agente GAIA...")
|
| 367 |
|
| 368 |
hf_token = os.getenv("HF_TOKEN")
|
| 369 |
+
gemini_api_key = os.getenv("GEMINI_API_KEY")
|
| 370 |
+
configured_model = os.getenv("GAIA_MODEL_ID")
|
| 371 |
+
|
| 372 |
+
model_id = configured_model or DEFAULT_HF_MODEL
|
| 373 |
+
|
| 374 |
+
if model_id.startswith("gemini/"):
|
| 375 |
+
model_api_key = gemini_api_key
|
| 376 |
+
required_secret = "GEMINI_API_KEY"
|
| 377 |
+
else:
|
| 378 |
+
model_api_key = hf_token
|
| 379 |
+
required_secret = "HF_TOKEN"
|
| 380 |
+
|
| 381 |
+
if not model_api_key:
|
| 382 |
raise RuntimeError(
|
| 383 |
+
f"O secret {required_secret} não está configurado. "
|
| 384 |
+
"Adicione a chave em Settings > Variables and secrets > Secrets."
|
|
|
|
| 385 |
)
|
| 386 |
|
| 387 |
self.model = LiteLLMModel(
|
| 388 |
+
model_id=model_id,
|
| 389 |
+
api_key=model_api_key,
|
| 390 |
temperature=0,
|
| 391 |
+
max_tokens=2_000,
|
| 392 |
)
|
| 393 |
self.hf_token = hf_token
|
| 394 |
+
self.model_id = model_id
|
| 395 |
+
print(f"Modelo principal selecionado: {model_id}")
|
| 396 |
+
agent_tools = [
|
| 397 |
+
DuckDuckGoSearchTool(max_results=8, rate_limit=1.0),
|
| 398 |
+
VisitWebpageTool(max_output_length=30_000),
|
| 399 |
+
WikipediaSearchTool(
|
| 400 |
+
user_agent="GAIA-Course-Agent/1.0 (educational project)",
|
| 401 |
+
language="en",
|
| 402 |
+
),
|
| 403 |
+
InspectGaiaAttachmentTool(),
|
| 404 |
+
YouTubeTranscriptTool(),
|
| 405 |
+
AnalyzeGaiaImageTool(),
|
| 406 |
+
]
|
| 407 |
+
if gemini_api_key:
|
| 408 |
+
agent_tools.append(ConsultGeminiTool())
|
| 409 |
+
|
| 410 |
self.agent = CodeAgent(
|
| 411 |
+
tools=agent_tools,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 412 |
model=self.model,
|
| 413 |
max_steps=10,
|
| 414 |
planning_interval=4,
|
|
|
|
| 438 |
with the exact video URL before searching the web.
|
| 439 |
If the task depends on an attached image, call analyze_gaia_image with the
|
| 440 |
task_id and complete question. Do not try to infer image contents from metadata.
|
| 441 |
+
When consult_gemini is available, use it selectively for a second opinion after
|
| 442 |
+
you have collected evidence, especially when sources conflict or your candidate
|
| 443 |
+
answer is uncertain. Give it the complete question and relevant evidence. Do
|
| 444 |
+
not blindly copy its response; compare it with the sources before deciding.
|
| 445 |
Prefer primary or official sources. When search snippets conflict, open the
|
| 446 |
source and verify the relevant passage instead of guessing.
|
| 447 |
Only call tools that are explicitly available. Never invent a function such as
|
|
|
|
| 479 |
or "Unauthorized" in error_text
|
| 480 |
):
|
| 481 |
raise RuntimeError(
|
| 482 |
+
f"Falha de autenticação no modelo {self.model_id}. "
|
| 483 |
+
"Verifique a chave secreta correspondente ao provedor "
|
| 484 |
+
"(GEMINI_API_KEY para Gemini ou HF_TOKEN para Hugging Face)."
|
| 485 |
) from exc
|
| 486 |
raise
|
| 487 |
return self.format_exact_answer(question, str(result))
|