Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -246,6 +246,9 @@
|
|
| 246 |
#
|
| 247 |
# =================================================================================================
|
| 248 |
|
|
|
|
|
|
|
|
|
|
| 249 |
import os
|
| 250 |
import io
|
| 251 |
import requests
|
|
@@ -258,42 +261,38 @@ import operator
|
|
| 258 |
# --- LangChain & LangGraph Imports ---
|
| 259 |
from langchain_core.messages import BaseMessage, HumanMessage, ToolMessage, AIMessage, SystemMessage
|
| 260 |
from langchain_core.tools import tool
|
| 261 |
-
|
|
|
|
| 262 |
from langgraph.graph import StateGraph, END
|
| 263 |
from langgraph.prebuilt import ToolNode
|
| 264 |
from tavily import TavilyClient
|
| 265 |
-
import pypdf
|
| 266 |
|
| 267 |
-
# (Keep Constants as is)
|
| 268 |
# --- Constants ---
|
| 269 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 270 |
FILES_DIR = "./files"
|
| 271 |
os.makedirs(FILES_DIR, exist_ok=True)
|
| 272 |
|
| 273 |
-
# --- System Prompt (Unchanged, it's
|
| 274 |
AGENT_SYSTEM_PROMPT = """You are a world-class AI agent, specialized in solving complex problems from the GAIA benchmark.
|
| 275 |
-
|
| 276 |
Your task is to analyze the user's question, think step-by-step, and use the provided tools to find the correct answer.
|
| 277 |
-
|
| 278 |
CRITICAL INSTRUCTIONS:
|
| 279 |
1. **Analyze the Goal:** First, understand what the user is asking for.
|
| 280 |
2. **Plan & Execute:** Formulate a plan and use the available tools (`tavily_search`, `read_file`, `python_interpreter`) to gather information.
|
| 281 |
3. **Final Answer Format:** Once you are absolutely certain of the answer, you MUST provide it directly and concisely.
|
| 282 |
- DO NOT include your reasoning, thoughts, or any conversational text like 'The answer is...', 'Here is the result:', or 'Based on my search...'.
|
| 283 |
- Your final response must ONLY be the answer itself.
|
| 284 |
-
|
| 285 |
EXAMPLES OF CORRECT FINAL ANSWERS:
|
| 286 |
- If the question asks for a year: `2023`
|
| 287 |
- If it asks for a name: `John Doe`
|
| 288 |
- If it asks for a number: `42`
|
| 289 |
- If it asks for a comma-separated list: `item1, item2, item3`
|
| 290 |
-
|
| 291 |
Think, use your tools, and then provide ONLY the final, precise answer.
|
| 292 |
"""
|
| 293 |
|
| 294 |
#
|
| 295 |
# ================================================================================================
|
| 296 |
-
# ✅ 1.
|
| 297 |
# ================================================================================================
|
| 298 |
#
|
| 299 |
tavily = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
|
|
@@ -326,7 +325,6 @@ def read_file(url: str) -> str:
|
|
| 326 |
with open(filename, 'wb') as f:
|
| 327 |
f.write(response.content)
|
| 328 |
|
| 329 |
-
# Check if the file is a PDF
|
| 330 |
if url.lower().endswith('.pdf'):
|
| 331 |
print(f"--- File identified as PDF. Reading with pypdf. ---")
|
| 332 |
try:
|
|
@@ -338,7 +336,6 @@ def read_file(url: str) -> str:
|
|
| 338 |
except Exception as e:
|
| 339 |
return f"Error reading PDF file: {e}"
|
| 340 |
else:
|
| 341 |
-
# Assume it's a text file
|
| 342 |
print(f"--- File identified as text. Reading normally. ---")
|
| 343 |
try:
|
| 344 |
with open(filename, 'r', encoding='utf-8') as f:
|
|
@@ -366,7 +363,7 @@ def python_interpreter(code: str) -> str:
|
|
| 366 |
|
| 367 |
#
|
| 368 |
# ================================================================================================
|
| 369 |
-
# ✅ 2. CONFIGURE AND BUILD THE AGENT GRAPH (NOW WITH
|
| 370 |
# ================================================================================================
|
| 371 |
#
|
| 372 |
class AgentState(TypedDict):
|
|
@@ -376,9 +373,15 @@ def build_agent_graph():
|
|
| 376 |
"""Builds the LangGraph agent."""
|
| 377 |
tools = [tavily_search, read_file, python_interpreter]
|
| 378 |
|
| 379 |
-
#
|
| 380 |
-
# It will use the
|
| 381 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 382 |
|
| 383 |
llm_with_tools = llm.bind_tools(tools)
|
| 384 |
|
|
@@ -406,7 +409,8 @@ def build_agent_graph():
|
|
| 406 |
#
|
| 407 |
class GaiaAgent:
|
| 408 |
def __init__(self):
|
| 409 |
-
print
|
|
|
|
| 410 |
self.agent_app = build_agent_graph()
|
| 411 |
|
| 412 |
def __call__(self, question: str) -> str:
|
|
@@ -418,7 +422,8 @@ class GaiaAgent:
|
|
| 418 |
]
|
| 419 |
}
|
| 420 |
final_state = None
|
| 421 |
-
|
|
|
|
| 422 |
if i == 0: print("--- Starting Agentic Loop ---")
|
| 423 |
final_state = step
|
| 424 |
|
|
@@ -427,6 +432,7 @@ class GaiaAgent:
|
|
| 427 |
print(f"\n--- Agent finished. Final Answer: {final_answer} ---\n")
|
| 428 |
return final_answer
|
| 429 |
|
|
|
|
| 430 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
| 431 |
space_id = os.getenv("SPACE_ID")
|
| 432 |
if not profile: return "Please Login to Hugging Face with the button.", None
|
|
@@ -470,14 +476,14 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
| 470 |
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
|
| 471 |
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 472 |
try:
|
| 473 |
-
response = requests.post(submit_url, json=submission_data, timeout=90)
|
| 474 |
response.raise_for_status()
|
| 475 |
result_data = response.json()
|
| 476 |
final_status = (
|
| 477 |
f"Submission Successful!\n"
|
| 478 |
f"User: {result_data.get('username')}\n"
|
| 479 |
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 480 |
-
f"({result_data.get('correct_count', '?')}/{
|
| 481 |
f"Message: {result_data.get('message', 'No message received.')}"
|
| 482 |
)
|
| 483 |
results_df = pd.DataFrame(results_log)
|
|
@@ -489,13 +495,14 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
| 489 |
|
| 490 |
# --- Gradio Interface ---
|
| 491 |
with gr.Blocks() as demo:
|
| 492 |
-
|
|
|
|
| 493 |
gr.Markdown(
|
| 494 |
"""
|
| 495 |
-
**Instructor's Note:** This
|
| 496 |
-
1. Ensure `
|
| 497 |
-
2. Ensure `requirements.txt` is updated
|
| 498 |
-
3.
|
| 499 |
"""
|
| 500 |
)
|
| 501 |
gr.LoginButton()
|
|
|
|
| 246 |
#
|
| 247 |
# =================================================================================================
|
| 248 |
|
| 249 |
+
#
|
| 250 |
+
|
| 251 |
+
##################
|
| 252 |
import os
|
| 253 |
import io
|
| 254 |
import requests
|
|
|
|
| 261 |
# --- LangChain & LangGraph Imports ---
|
| 262 |
from langchain_core.messages import BaseMessage, HumanMessage, ToolMessage, AIMessage, SystemMessage
|
| 263 |
from langchain_core.tools import tool
|
| 264 |
+
# <<<--- CHANGE 1: Import Google Gemini instead of OpenAI --->>>
|
| 265 |
+
from langchain_google_genai import ChatGoogleGenerativeAI
|
| 266 |
from langgraph.graph import StateGraph, END
|
| 267 |
from langgraph.prebuilt import ToolNode
|
| 268 |
from tavily import TavilyClient
|
| 269 |
+
import pypdf
|
| 270 |
|
|
|
|
| 271 |
# --- Constants ---
|
| 272 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 273 |
FILES_DIR = "./files"
|
| 274 |
os.makedirs(FILES_DIR, exist_ok=True)
|
| 275 |
|
| 276 |
+
# --- System Prompt (Unchanged, it's strong) ---
|
| 277 |
AGENT_SYSTEM_PROMPT = """You are a world-class AI agent, specialized in solving complex problems from the GAIA benchmark.
|
|
|
|
| 278 |
Your task is to analyze the user's question, think step-by-step, and use the provided tools to find the correct answer.
|
|
|
|
| 279 |
CRITICAL INSTRUCTIONS:
|
| 280 |
1. **Analyze the Goal:** First, understand what the user is asking for.
|
| 281 |
2. **Plan & Execute:** Formulate a plan and use the available tools (`tavily_search`, `read_file`, `python_interpreter`) to gather information.
|
| 282 |
3. **Final Answer Format:** Once you are absolutely certain of the answer, you MUST provide it directly and concisely.
|
| 283 |
- DO NOT include your reasoning, thoughts, or any conversational text like 'The answer is...', 'Here is the result:', or 'Based on my search...'.
|
| 284 |
- Your final response must ONLY be the answer itself.
|
|
|
|
| 285 |
EXAMPLES OF CORRECT FINAL ANSWERS:
|
| 286 |
- If the question asks for a year: `2023`
|
| 287 |
- If it asks for a name: `John Doe`
|
| 288 |
- If it asks for a number: `42`
|
| 289 |
- If it asks for a comma-separated list: `item1, item2, item3`
|
|
|
|
| 290 |
Think, use your tools, and then provide ONLY the final, precise answer.
|
| 291 |
"""
|
| 292 |
|
| 293 |
#
|
| 294 |
# ================================================================================================
|
| 295 |
+
# ✅ 1. AGENT'S TOOLS (Unchanged)
|
| 296 |
# ================================================================================================
|
| 297 |
#
|
| 298 |
tavily = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
|
|
|
|
| 325 |
with open(filename, 'wb') as f:
|
| 326 |
f.write(response.content)
|
| 327 |
|
|
|
|
| 328 |
if url.lower().endswith('.pdf'):
|
| 329 |
print(f"--- File identified as PDF. Reading with pypdf. ---")
|
| 330 |
try:
|
|
|
|
| 336 |
except Exception as e:
|
| 337 |
return f"Error reading PDF file: {e}"
|
| 338 |
else:
|
|
|
|
| 339 |
print(f"--- File identified as text. Reading normally. ---")
|
| 340 |
try:
|
| 341 |
with open(filename, 'r', encoding='utf-8') as f:
|
|
|
|
| 363 |
|
| 364 |
#
|
| 365 |
# ================================================================================================
|
| 366 |
+
# ✅ 2. CONFIGURE AND BUILD THE AGENT GRAPH (NOW WITH GEMINI 1.5 PRO)
|
| 367 |
# ================================================================================================
|
| 368 |
#
|
| 369 |
class AgentState(TypedDict):
|
|
|
|
| 373 |
"""Builds the LangGraph agent."""
|
| 374 |
tools = [tavily_search, read_file, python_interpreter]
|
| 375 |
|
| 376 |
+
# <<<--- CHANGE 2: Instantiate the Gemini Model --->>>
|
| 377 |
+
# It will use the GOOGLE_API_KEY from your secrets.
|
| 378 |
+
# Note: `convert_system_message_to_human=True` is a useful flag for compatibility,
|
| 379 |
+
# ensuring our powerful system prompt is always understood correctly by the Gemini model.
|
| 380 |
+
llm = ChatGoogleGenerativeAI(
|
| 381 |
+
model="gemini-1.5-pro-latest",
|
| 382 |
+
temperature=0,
|
| 383 |
+
convert_system_message_to_human=True
|
| 384 |
+
)
|
| 385 |
|
| 386 |
llm_with_tools = llm.bind_tools(tools)
|
| 387 |
|
|
|
|
| 409 |
#
|
| 410 |
class GaiaAgent:
|
| 411 |
def __init__(self):
|
| 412 |
+
# <<<--- CHANGE 3: Update the print statement for clarity --->>>
|
| 413 |
+
print("GaiaAgent initialized. Building fresh Gemini 1.5 Pro graph...")
|
| 414 |
self.agent_app = build_agent_graph()
|
| 415 |
|
| 416 |
def __call__(self, question: str) -> str:
|
|
|
|
| 422 |
]
|
| 423 |
}
|
| 424 |
final_state = None
|
| 425 |
+
# We increase the recursion limit slightly, as some models might take an extra step for reasoning.
|
| 426 |
+
for i, step in enumerate(self.agent_app.stream(initial_input, {"recursion_limit": 20})):
|
| 427 |
if i == 0: print("--- Starting Agentic Loop ---")
|
| 428 |
final_state = step
|
| 429 |
|
|
|
|
| 432 |
print(f"\n--- Agent finished. Final Answer: {final_answer} ---\n")
|
| 433 |
return final_answer
|
| 434 |
|
| 435 |
+
# (The rest of the file remains exactly the same)
|
| 436 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
| 437 |
space_id = os.getenv("SPACE_ID")
|
| 438 |
if not profile: return "Please Login to Hugging Face with the button.", None
|
|
|
|
| 476 |
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
|
| 477 |
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 478 |
try:
|
| 479 |
+
response = requests.post(submit_url, json=submission_data, timeout=90)
|
| 480 |
response.raise_for_status()
|
| 481 |
result_data = response.json()
|
| 482 |
final_status = (
|
| 483 |
f"Submission Successful!\n"
|
| 484 |
f"User: {result_data.get('username')}\n"
|
| 485 |
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 486 |
+
f"({result_data.get('correct_count', '?')}/{result_a.get('total_attempted', '?')} correct)\n"
|
| 487 |
f"Message: {result_data.get('message', 'No message received.')}"
|
| 488 |
)
|
| 489 |
results_df = pd.DataFrame(results_log)
|
|
|
|
| 495 |
|
| 496 |
# --- Gradio Interface ---
|
| 497 |
with gr.Blocks() as demo:
|
| 498 |
+
# <<<--- CHANGE 4: Update the title in the UI --->>>
|
| 499 |
+
gr.Markdown("# GAIA Agent Final Assessment (V6 - Gemini 1.5 Pro)")
|
| 500 |
gr.Markdown(
|
| 501 |
"""
|
| 502 |
+
**Instructor's Note:** This version has been upgraded to use Google's `gemini-1.5-pro-latest` model.
|
| 503 |
+
1. Ensure `GOOGLE_API_KEY` and `TAVILY_API_KEY` are set in your Space secrets.
|
| 504 |
+
2. Ensure `requirements.txt` is updated with `langchain-google-genai`.
|
| 505 |
+
3. Let's see how Gemini performs! Good luck.
|
| 506 |
"""
|
| 507 |
)
|
| 508 |
gr.LoginButton()
|