Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -199,7 +199,6 @@
|
|
| 199 |
# demo.launch(debug=True, share=False)
|
| 200 |
|
| 201 |
|
| 202 |
-
|
| 203 |
import os
|
| 204 |
import gradio as gr
|
| 205 |
import requests
|
|
@@ -210,44 +209,41 @@ from typing import TypedDict, Annotated, List
|
|
| 210 |
|
| 211 |
# ==============================================================================
|
| 212 |
# PART 1: YOUR AGENT'S LOGIC AND DEFINITION
|
| 213 |
-
# All of this is new. It replaces the old placeholder.
|
| 214 |
# ==============================================================================
|
| 215 |
|
| 216 |
# LangChain and LangGraph imports
|
| 217 |
from langchain_huggingface import HuggingFaceEndpoint
|
| 218 |
-
from
|
|
|
|
| 219 |
from langchain_experimental.tools import PythonREPLTool
|
| 220 |
from langchain_core.messages import BaseMessage, HumanMessage
|
| 221 |
from langgraph.graph import StateGraph, END
|
| 222 |
from langgraph.prebuilt import ToolNode
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
|
| 224 |
# Load API keys from .env file or Space secrets
|
| 225 |
load_dotenv()
|
| 226 |
hf_token = os.getenv("HF_TOKEN")
|
| 227 |
tavily_api_key = os.getenv("TAVILY_API_KEY")
|
| 228 |
|
| 229 |
-
# Set the Tavily API key for the tool to use
|
| 230 |
if tavily_api_key:
|
| 231 |
os.environ["TAVILY_API_KEY"] = tavily_api_key
|
| 232 |
else:
|
| 233 |
print("Warning: TAVILY_API_KEY not found. Web search tool will not work.")
|
| 234 |
|
| 235 |
# --- Define Agent Tools ---
|
|
|
|
| 236 |
tools = [
|
| 237 |
-
|
| 238 |
PythonREPLTool()
|
| 239 |
]
|
| 240 |
tool_node = ToolNode(tools)
|
| 241 |
|
| 242 |
# --- Configure the LLM "Brain" ---
|
| 243 |
repo_id = "meta-llama/Meta-Llama-3-8B-Instruct"
|
| 244 |
-
SYSTEM_PROMPT = """You are a highly capable AI agent. Your mission is to accurately answer complex questions.
|
| 245 |
-
**Instructions:**
|
| 246 |
-
1. **Analyze:** Read the question to understand what is being asked.
|
| 247 |
-
2. **Plan:** Think step-by-step. Break the problem into smaller tasks. Decide which tool is best for each task.
|
| 248 |
-
3. **Execute:** Call ONE tool at a time.
|
| 249 |
-
4. **Observe & Reason:** After getting a tool's result, observe it. Decide if you have the final answer or if you need to use another tool.
|
| 250 |
-
5. **Final Answer:** Once confident, provide a clear, direct, and concise final answer."""
|
| 251 |
|
| 252 |
llm = HuggingFaceEndpoint(
|
| 253 |
repo_id=repo_id,
|
|
@@ -255,179 +251,130 @@ llm = HuggingFaceEndpoint(
|
|
| 255 |
temperature=0,
|
| 256 |
max_new_tokens=2048,
|
| 257 |
)
|
| 258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 259 |
|
| 260 |
# --- Build the LangGraph Agent ---
|
| 261 |
class AgentState(TypedDict):
|
| 262 |
-
messages
|
|
|
|
|
|
|
|
|
|
| 263 |
|
|
|
|
| 264 |
def agent_node(state):
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
# Compile the graph into a runnable app
|
| 283 |
-
compiled_agent_app = workflow.compile()
|
| 284 |
-
|
| 285 |
-
# --- THIS IS THE NEW BasicAgent CLASS THAT REPLACES THE PLACEHOLDER ---
|
| 286 |
class BasicAgent:
|
| 287 |
def __init__(self):
|
| 288 |
-
# Check for API keys during initialization
|
| 289 |
if not hf_token or not tavily_api_key:
|
| 290 |
raise ValueError("HF_TOKEN or TAVILY_API_KEY not set. Please add them to your Space secrets.")
|
| 291 |
print("LangGraph Agent initialized successfully.")
|
|
|
|
|
|
|
|
|
|
| 292 |
|
| 293 |
def __call__(self, question: str) -> str:
|
| 294 |
print(f"Agent received question (first 80 chars): {question[:80]}...")
|
| 295 |
try:
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
if not final_response:
|
| 303 |
-
final_response = "Agent finished but did not produce a clear final answer."
|
| 304 |
-
|
| 305 |
-
print(f"Agent returning final answer (first 80 chars): {final_response[:80]}...")
|
| 306 |
-
return final_response
|
| 307 |
except Exception as e:
|
| 308 |
print(f"An error occurred in agent execution: {e}")
|
| 309 |
return f"Error: {e}"
|
| 310 |
|
|
|
|
| 311 |
# ==============================================================================
|
| 312 |
-
# PART 2: THE GRADIO TEST HARNESS UI
|
| 313 |
-
# The rest of this file remains exactly as it was provided in the template.
|
| 314 |
# ==============================================================================
|
| 315 |
-
|
| 316 |
# --- Constants ---
|
| 317 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 318 |
|
| 319 |
def run_and_submit_all(profile: gr.OAuthProfile | None):
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
and displays the results.
|
| 323 |
-
"""
|
| 324 |
-
# (The rest of this function remains unchanged)
|
| 325 |
-
# --- Determine HF Space Runtime URL and Repo URL ---
|
| 326 |
-
space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code
|
| 327 |
-
|
| 328 |
if profile:
|
| 329 |
username= f"{profile.username}"
|
| 330 |
print(f"User logged in: {username}")
|
| 331 |
else:
|
| 332 |
print("User not logged in.")
|
| 333 |
return "Please Login to Hugging Face with the button.", None
|
| 334 |
-
|
| 335 |
api_url = DEFAULT_API_URL
|
| 336 |
questions_url = f"{api_url}/questions"
|
| 337 |
submit_url = f"{api_url}/submit"
|
| 338 |
-
|
| 339 |
-
# 1. Instantiate Agent (this now calls YOUR agent class from above)
|
| 340 |
try:
|
| 341 |
agent = BasicAgent()
|
| 342 |
except Exception as e:
|
| 343 |
print(f"Error instantiating agent: {e}")
|
| 344 |
return f"Error initializing agent: {e}", None
|
| 345 |
-
|
| 346 |
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
|
| 347 |
-
print(agent_code)
|
| 348 |
-
|
| 349 |
-
# 2. Fetch Questions
|
| 350 |
print(f"Fetching questions from: {questions_url}")
|
| 351 |
try:
|
| 352 |
response = requests.get(questions_url, timeout=15)
|
| 353 |
response.raise_for_status()
|
| 354 |
questions_data = response.json()
|
| 355 |
-
if not questions_data:
|
| 356 |
-
print("Fetched questions list is empty.")
|
| 357 |
-
return "Fetched questions list is empty or invalid format.", None
|
| 358 |
print(f"Fetched {len(questions_data)} questions.")
|
| 359 |
except Exception as e:
|
| 360 |
-
print(f"An unexpected error occurred fetching questions: {e}")
|
| 361 |
return f"An unexpected error occurred fetching questions: {e}", None
|
| 362 |
-
|
| 363 |
-
# 3. Run your Agent
|
| 364 |
-
results_log = []
|
| 365 |
-
answers_payload = []
|
| 366 |
print(f"Running agent on {len(questions_data)} questions...")
|
| 367 |
for item in questions_data:
|
| 368 |
-
task_id = item.get("task_id")
|
| 369 |
-
question_text
|
| 370 |
-
if not task_id or question_text is None:
|
| 371 |
-
print(f"Skipping item with missing task_id or question: {item}")
|
| 372 |
-
continue
|
| 373 |
try:
|
| 374 |
-
# This line now calls your __call__ method in your new BasicAgent
|
| 375 |
submitted_answer = agent(question_text)
|
| 376 |
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
|
| 377 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
| 378 |
except Exception as e:
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
if not answers_payload:
|
| 383 |
-
print("Agent did not produce any answers to submit.")
|
| 384 |
-
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|
| 385 |
-
|
| 386 |
-
# 4. Prepare Submission
|
| 387 |
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
|
| 388 |
-
status_update = f"Agent finished. Submitting {len(answers_payload)} answers for user '{username}'..."
|
| 389 |
-
print(status_update)
|
| 390 |
-
|
| 391 |
-
# 5. Submit
|
| 392 |
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 393 |
try:
|
| 394 |
response = requests.post(submit_url, json=submission_data, timeout=60)
|
| 395 |
response.raise_for_status()
|
| 396 |
result_data = response.json()
|
| 397 |
-
final_status = (
|
| 398 |
-
|
| 399 |
-
f"User: {result_data.get('username')}\n"
|
| 400 |
-
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 401 |
-
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
|
| 402 |
-
f"Message: {result_data.get('message', 'No message received.')}"
|
| 403 |
-
)
|
| 404 |
-
print("Submission successful.")
|
| 405 |
-
results_df = pd.DataFrame(results_log)
|
| 406 |
-
return final_status, results_df
|
| 407 |
except Exception as e:
|
| 408 |
-
|
| 409 |
-
print(status_message)
|
| 410 |
-
results_df = pd.DataFrame(results_log)
|
| 411 |
-
return status_message, results_df
|
| 412 |
-
|
| 413 |
|
| 414 |
-
# ---
|
| 415 |
with gr.Blocks() as demo:
|
| 416 |
gr.Markdown("# GAIA Agent Evaluation Runner")
|
| 417 |
-
gr.Markdown(
|
| 418 |
-
"""
|
| 419 |
-
1. Log in to your Hugging Face account using the button below. This uses your HF username for submission.
|
| 420 |
-
2. Click 'Run Evaluation & Submit All Answers' to run your custom agent and see the score.
|
| 421 |
-
"""
|
| 422 |
-
)
|
| 423 |
gr.LoginButton()
|
| 424 |
run_button = gr.Button("Run Evaluation & Submit All Answers")
|
| 425 |
status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
|
| 426 |
results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
|
| 427 |
-
run_button.click(
|
| 428 |
-
fn=run_and_submit_all,
|
| 429 |
-
outputs=[status_output, results_table]
|
| 430 |
-
)
|
| 431 |
|
| 432 |
if __name__ == "__main__":
|
| 433 |
demo.launch(debug=True, share=False)
|
|
|
|
| 199 |
# demo.launch(debug=True, share=False)
|
| 200 |
|
| 201 |
|
|
|
|
| 202 |
import os
|
| 203 |
import gradio as gr
|
| 204 |
import requests
|
|
|
|
| 209 |
|
| 210 |
# ==============================================================================
|
| 211 |
# PART 1: YOUR AGENT'S LOGIC AND DEFINITION
|
|
|
|
| 212 |
# ==============================================================================
|
| 213 |
|
| 214 |
# LangChain and LangGraph imports
|
| 215 |
from langchain_huggingface import HuggingFaceEndpoint
|
| 216 |
+
# NEW: Import TavilySearch from the new package
|
| 217 |
+
from langchain_tavily import TavilySearch
|
| 218 |
from langchain_experimental.tools import PythonREPLTool
|
| 219 |
from langchain_core.messages import BaseMessage, HumanMessage
|
| 220 |
from langgraph.graph import StateGraph, END
|
| 221 |
from langgraph.prebuilt import ToolNode
|
| 222 |
+
# NEW: Import the compatible agent constructor and prompt hub
|
| 223 |
+
from langchain.agents import create_tool_calling_agent
|
| 224 |
+
from langchain import hub
|
| 225 |
+
|
| 226 |
|
| 227 |
# Load API keys from .env file or Space secrets
|
| 228 |
load_dotenv()
|
| 229 |
hf_token = os.getenv("HF_TOKEN")
|
| 230 |
tavily_api_key = os.getenv("TAVILY_API_KEY")
|
| 231 |
|
|
|
|
| 232 |
if tavily_api_key:
|
| 233 |
os.environ["TAVILY_API_KEY"] = tavily_api_key
|
| 234 |
else:
|
| 235 |
print("Warning: TAVILY_API_KEY not found. Web search tool will not work.")
|
| 236 |
|
| 237 |
# --- Define Agent Tools ---
|
| 238 |
+
# NEW: Using TavilySearch from the correct package
|
| 239 |
tools = [
|
| 240 |
+
TavilySearch(max_results=3, description="A search engine for finding up-to-date information on the web."),
|
| 241 |
PythonREPLTool()
|
| 242 |
]
|
| 243 |
tool_node = ToolNode(tools)
|
| 244 |
|
| 245 |
# --- Configure the LLM "Brain" ---
|
| 246 |
repo_id = "meta-llama/Meta-Llama-3-8B-Instruct"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
|
| 248 |
llm = HuggingFaceEndpoint(
|
| 249 |
repo_id=repo_id,
|
|
|
|
| 251 |
temperature=0,
|
| 252 |
max_new_tokens=2048,
|
| 253 |
)
|
| 254 |
+
|
| 255 |
+
# --- THE FIX: Create Agent with a Compatible Method ---
|
| 256 |
+
# REMOVED: llm_with_tools = llm.bind_tools(tools)
|
| 257 |
+
# This was causing the error.
|
| 258 |
+
|
| 259 |
+
# NEW: We pull a pre-made prompt that knows how to handle tool calls.
|
| 260 |
+
prompt = hub.pull("hwchase17/react-json")
|
| 261 |
+
|
| 262 |
+
# NEW: We use `create_tool_calling_agent`. This function correctly combines the LLM,
|
| 263 |
+
# the tools, and the prompt, without needing the .bind_tools() method.
|
| 264 |
+
agent_runnable = create_tool_calling_agent(llm, tools, prompt)
|
| 265 |
+
|
| 266 |
|
| 267 |
# --- Build the LangGraph Agent ---
|
| 268 |
class AgentState(TypedDict):
|
| 269 |
+
# The 'messages' key is no longer used, 'input' and 'agent_outcome' are standard for this agent type
|
| 270 |
+
input: str
|
| 271 |
+
chat_history: list[BaseMessage]
|
| 272 |
+
agent_outcome: dict
|
| 273 |
|
| 274 |
+
# NEW: The agent_node is much simpler now. It just calls the runnable we created.
|
| 275 |
def agent_node(state):
|
| 276 |
+
outcome = agent_runnable.invoke(state)
|
| 277 |
+
return {"agent_outcome": outcome}
|
| 278 |
+
|
| 279 |
+
def tool_node_executor(state):
|
| 280 |
+
# The agent_runnable provides tool calls in a specific format. We execute them.
|
| 281 |
+
tool_calls = state["agent_outcome"].tool_calls
|
| 282 |
+
tool_outputs = []
|
| 283 |
+
for tool_call in tool_calls:
|
| 284 |
+
tool_name = tool_call["name"]
|
| 285 |
+
tool_to_call = {tool.name: tool for tool in tools}[tool_name]
|
| 286 |
+
tool_output = tool_to_call.invoke(tool_call["args"])
|
| 287 |
+
tool_outputs.append({"output": tool_output, "tool_call_id": tool_call["id"]})
|
| 288 |
+
return {"intermediate_steps": tool_outputs}
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
# This setup is more complex but correctly models the ReAct loop in LangGraph
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 292 |
class BasicAgent:
|
| 293 |
def __init__(self):
|
|
|
|
| 294 |
if not hf_token or not tavily_api_key:
|
| 295 |
raise ValueError("HF_TOKEN or TAVILY_API_KEY not set. Please add them to your Space secrets.")
|
| 296 |
print("LangGraph Agent initialized successfully.")
|
| 297 |
+
# We need an agent executor to run the loop
|
| 298 |
+
from langchain.agents import AgentExecutor
|
| 299 |
+
self.agent_executor = AgentExecutor(agent=agent_runnable, tools=tools, verbose=True)
|
| 300 |
|
| 301 |
def __call__(self, question: str) -> str:
|
| 302 |
print(f"Agent received question (first 80 chars): {question[:80]}...")
|
| 303 |
try:
|
| 304 |
+
# The AgentExecutor expects a dictionary with an "input" key.
|
| 305 |
+
response = self.agent_executor.invoke({"input": question})
|
| 306 |
+
final_answer = response.get("output", "Agent did not produce an output.")
|
| 307 |
+
print(f"Agent returning final answer (first 80 chars): {final_answer[:80]}...")
|
| 308 |
+
return final_answer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 309 |
except Exception as e:
|
| 310 |
print(f"An error occurred in agent execution: {e}")
|
| 311 |
return f"Error: {e}"
|
| 312 |
|
| 313 |
+
|
| 314 |
# ==============================================================================
|
| 315 |
+
# PART 2: THE GRADIO TEST HARNESS UI (UNCHANGED)
|
|
|
|
| 316 |
# ==============================================================================
|
|
|
|
| 317 |
# --- Constants ---
|
| 318 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 319 |
|
| 320 |
def run_and_submit_all(profile: gr.OAuthProfile | None):
|
| 321 |
+
# This entire function remains the same as the template
|
| 322 |
+
space_id = os.getenv("SPACE_ID")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 323 |
if profile:
|
| 324 |
username= f"{profile.username}"
|
| 325 |
print(f"User logged in: {username}")
|
| 326 |
else:
|
| 327 |
print("User not logged in.")
|
| 328 |
return "Please Login to Hugging Face with the button.", None
|
|
|
|
| 329 |
api_url = DEFAULT_API_URL
|
| 330 |
questions_url = f"{api_url}/questions"
|
| 331 |
submit_url = f"{api_url}/submit"
|
|
|
|
|
|
|
| 332 |
try:
|
| 333 |
agent = BasicAgent()
|
| 334 |
except Exception as e:
|
| 335 |
print(f"Error instantiating agent: {e}")
|
| 336 |
return f"Error initializing agent: {e}", None
|
|
|
|
| 337 |
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
|
|
|
|
|
|
|
|
|
|
| 338 |
print(f"Fetching questions from: {questions_url}")
|
| 339 |
try:
|
| 340 |
response = requests.get(questions_url, timeout=15)
|
| 341 |
response.raise_for_status()
|
| 342 |
questions_data = response.json()
|
|
|
|
|
|
|
|
|
|
| 343 |
print(f"Fetched {len(questions_data)} questions.")
|
| 344 |
except Exception as e:
|
|
|
|
| 345 |
return f"An unexpected error occurred fetching questions: {e}", None
|
| 346 |
+
results_log, answers_payload = [], []
|
|
|
|
|
|
|
|
|
|
| 347 |
print(f"Running agent on {len(questions_data)} questions...")
|
| 348 |
for item in questions_data:
|
| 349 |
+
task_id, question_text = item.get("task_id"), item.get("question")
|
| 350 |
+
if not task_id or question_text is None: continue
|
|
|
|
|
|
|
|
|
|
| 351 |
try:
|
|
|
|
| 352 |
submitted_answer = agent(question_text)
|
| 353 |
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
|
| 354 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
| 355 |
except Exception as e:
|
| 356 |
+
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
| 357 |
+
if not answers_payload: return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 358 |
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 360 |
try:
|
| 361 |
response = requests.post(submit_url, json=submission_data, timeout=60)
|
| 362 |
response.raise_for_status()
|
| 363 |
result_data = response.json()
|
| 364 |
+
final_status = (f"Submission Successful!\nUser: {result_data.get('username')}\nOverall Score: {result_data.get('score', 'N/A')}% ({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\nMessage: {result_data.get('message', '')}")
|
| 365 |
+
return final_status, pd.DataFrame(results_log)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 366 |
except Exception as e:
|
| 367 |
+
return f"An unexpected error occurred during submission: {e}", pd.DataFrame(results_log)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 368 |
|
| 369 |
+
# --- Gradio Interface (Unchanged) ---
|
| 370 |
with gr.Blocks() as demo:
|
| 371 |
gr.Markdown("# GAIA Agent Evaluation Runner")
|
| 372 |
+
gr.Markdown("1. Log in. 2. Click 'Run Evaluation'.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 373 |
gr.LoginButton()
|
| 374 |
run_button = gr.Button("Run Evaluation & Submit All Answers")
|
| 375 |
status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
|
| 376 |
results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
|
| 377 |
+
run_button.click(fn=run_and_submit_all, outputs=[status_output, results_table])
|
|
|
|
|
|
|
|
|
|
| 378 |
|
| 379 |
if __name__ == "__main__":
|
| 380 |
demo.launch(debug=True, share=False)
|