Mouhamedamar's picture
Update app.py
6076917 verified
import os
import gradio as gr
import requests
import pandas as pd
import re
from urllib.parse import urlparse
from typing import TypedDict, List, Optional, Annotated, Tuple, Union, Literal
from langgraph.graph import StateGraph, END
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, ToolMessage, BaseMessage
from langgraph.graph.message import add_messages
from langchain_core.tools import tool
from langchain_community.tools import WikipediaQueryRun
from langchain_community.utilities import WikipediaAPIWrapper
from langchain_tavily import TavilySearch
from pydantic import BaseModel, Field
from langgraph.prebuilt import ToolNode
from langchain_core.prompts import ChatPromptTemplate
import operator
# --- Constants ---
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
TEMP_DIR_BASE = os.path.join(os.getcwd(), "temp_agent_files")
# --- Helper Functions ---
def get_task_temp_dir(task_id: str) -> str:
"""Creates and returns a unique temporary directory for a task."""
task_dir = os.path.join(TEMP_DIR_BASE, task_id)
os.makedirs(task_dir, exist_ok=True)
return task_dir
def extract_youtube_id(url: str) -> Optional[str]:
"""Extract YouTube video ID from URL."""
pattern = r'(?:youtube\.com\/(?:watch\?v=|embed\/)|youtu\.be\/)([a-zA-Z0-9_-]+)'
match = re.search(pattern, url)
return match.group(1) if match else None
# --- Analysis Tools with Gemini ---
@tool
def analyze_youtube_video(url: str, question: str) -> str:
"""
Analyze a YouTube video using Gemini 2.0 Flash Thinking.
Args:
url: The YouTube video URL
question: Specific question about the video content
Returns:
Analysis of the video based on the provided question.
"""
try:
parsed_url = urlparse(url)
if not all([parsed_url.scheme, parsed_url.netloc]):
return "Please provide a valid video URL with http:// or https:// prefix."
if 'youtube.com' not in url and 'youtu.be' not in url:
return "Only YouTube videos are supported at this time."
api_key = os.environ.get("GOOGLE_API_KEY")
if not api_key:
return "Unable to perform analysis: Google API key not set. Get it from https://aistudio.google.com/"
llm = ChatGoogleGenerativeAI(
model="gemini-2.5-flash",
google_api_key=api_key,
temperature=0,
max_output_tokens=4096
)
prompt = f"""You are analyzing a YouTube video at URL: {url}
Question about the video: {question}
Based on what you know about this video (if it's a known video) or general knowledge,
provide a helpful analysis. If you cannot access the video directly, provide
reasonable information based on the video title/URL if it's recognizable.
Analysis:"""
response = llm.invoke(prompt)
return f"## YouTube Video Analysis (URL: {url})\n\n{response.content}"
except Exception as e:
print(f"Error in analyze_youtube_video: {type(e).__name__}: {e}")
return f"Error analyzing video at {url}: {str(e)}"
@tool
def analyze_text_content(content: str, question: str) -> str:
"""
Analyze text content using Gemini.
Args:
content: The text content to analyze
question: Specific question about the content
Returns:
Analysis of the text based on the question.
"""
try:
api_key = os.environ.get("GOOGLE_API_KEY")
if not api_key:
return "Unable to perform analysis: Google API key not set."
llm = ChatGoogleGenerativeAI(
model="gemini-2.5-flash",
google_api_key=api_key,
temperature=0,
max_output_tokens=4096
)
prompt = f"""Analyze the following content and answer the question.
Content: {content[:8000]}
Question: {question}
Provide a concise, accurate answer based ONLY on the content above.
If the content doesn't contain the answer, say "Information not found in the provided content."
Answer:"""
response = llm.invoke(prompt)
return response.content
except Exception as e:
return f"Error analyzing text: {str(e)}"
@tool
def direct_reasoning(question: str, context: str = "") -> str:
"""
Use Gemini's reasoning capabilities to answer a question.
Args:
question: The question to answer
context: Optional context to help answer
Returns:
The reasoned answer
"""
try:
api_key = os.environ.get("GOOGLE_API_KEY")
if not api_key:
return "Google API key not set."
llm = ChatGoogleGenerativeAI(
model="gemini-2.5-flash",
google_api_key=api_key,
temperature=0,
max_output_tokens=4096
)
prompt = f"""Answer the following question with ONLY the exact answer, nothing else.
No explanations, no "FINAL ANSWER", just the answer.
{context}
Question: {question}
Answer:"""
response = llm.invoke(prompt)
return response.content.strip()
except Exception as e:
return f"Error: {str(e)}"
# --- Agent State ---
class TaskState(TypedDict):
task_id: str
question: str
file_name: Optional[str]
api_url: str
file_path: Optional[str]
temp_dir: Optional[str]
plan: List[str]
past_steps: Annotated[List[Tuple[str, str]], operator.add]
response: str
messages: Annotated[list[BaseMessage], add_messages]
current_task: str
# --- Search Tool Setup ---
def setup_tavily_search():
"""Set up Tavily search tool"""
try:
tavily_api_key = os.environ.get("TAVILY_API_KEY")
if not tavily_api_key:
raise ValueError("Tavily API key not found. Set TAVILY_API_KEY environment variable.")
print("Using Tavily for web search")
return TavilySearch(max_results=10)
except Exception as e:
print(f"Error setting up Tavily: {e}")
raise
# --- LLM Initialization with Gemini ---
def get_llm():
"""Get Gemini LLM instance"""
api_key = os.environ.get("GOOGLE_API_KEY")
if not api_key:
raise ValueError("GOOGLE_API_KEY environment variable not set. Get it from https://aistudio.google.com/")
return ChatGoogleGenerativeAI(
model="gemini-2.5-flash",
google_api_key=api_key,
temperature=0,
max_output_tokens=4096
)
llm = get_llm()
# --- Tool Definitions ---
web_search = setup_tavily_search()
wikipedia_api = WikipediaAPIWrapper(top_k_results=8, use_https=True)
wikipedia_search = WikipediaQueryRun(api_wrapper=wikipedia_api)
tools = [
analyze_youtube_video,
analyze_text_content,
direct_reasoning,
web_search,
wikipedia_search
]
tool_node = ToolNode(tools)
# --- Pydantic Models for Planner/Replanner ---
class Plan(BaseModel):
"""Plan to follow in future"""
thought: str = Field(description="The reasoning process behind generating this plan.")
steps: List[str] = Field(description="Different steps to follow, in sorted order.")
class Response(BaseModel):
"""Response to user."""
response: str
class Act(BaseModel):
"""Action to perform."""
thought: str = Field(description="The reasoning process behind choosing this action (Plan or Response).")
action: Union[Response, Plan] = Field(description="Action to perform. Response for final answer, Plan for more steps.")
# --- Planner Prompt Setup ---
def get_tools_description() -> str:
"""Generate a formatted string describing all available tools."""
tool_descriptions = []
for tool in tools:
name = getattr(tool, "name", str(tool))
description = getattr(tool, "description", getattr(tool, "__doc__", "No description available"))
first_line_desc = description.split('\n')[0].strip() if description else "No description available"
tool_descriptions.append(f"- `{name}`: {first_line_desc}")
return "\n".join(tool_descriptions)
tools_desc = get_tools_description()
planner_prompt = ChatPromptTemplate.from_messages(
[
(
"system",
f"""For the given objective, devise a simple step-by-step plan.
Also provide a detailed thought process explaining how you arrived at the plan.
**Plan Requirements:**
* **Simplicity:** Keep the plan as straightforward as possible.
* **Task Types:** Each step must be EITHER:
* A task requiring a specific tool from the available list.
* A reasoning step for the LLM to perform internally (e.g., summarizing information, comparing results).
* **Tool Usage:** If a step uses a tool, clearly state the tool name and what it should do.
* **Conciseness:** Avoid superfluous steps. The result of the final step should be the final answer.
**Available Tools:**
{tools_desc}
Output your thought process and the plan steps.
""",
),
("placeholder", "{initial_user_message}"),
]
)
planner = planner_prompt | llm.with_structured_output(Plan)
# --- Replanner Prompt Setup ---
replanner_prompt = ChatPromptTemplate.from_template(
f"""You are a replanner. Your goal is to refine the plan to achieve the objective, or decide if the objective is met.
**Objective:**
{{question}}
**Original Plan (remaining steps):**
{{plan_str}}
**History (Executed Steps and Thoughts):**
{{past_steps_str}}
**Most Recent Step Executed:** '{{current_task}}'
**Direct Result of Last Step:**
{{latest_result}}
**Your Task:**
Analyze the **History (Executed Steps and Thoughts)** and the **Direct Result of Last Step** carefully.
* If the last step successfully moved towards the objective, continue the plan or refine it.
* If the last step failed, resulted in an error, or the **History** suggests the current approach is not working, you MUST revise the plan to try a different approach.
Based on this analysis, decide the next course of action (Respond or Revise Plan).
**Action Options:**
1. **Respond (Response action):** If the objective is met and you have the final answer, provide it.
2. **Revise Plan (Plan action):** If more steps are needed, provide a new, simple plan containing only the remaining steps.
**Available Tools:**
{tools_desc}
Output your thought process and the chosen action (Plan or Response).
"""
)
replanner = replanner_prompt | llm.with_structured_output(Act)
# --- Agent Node Functions ---
def plan_step(state: TaskState):
"""Generate the initial plan based on the initial question/file info."""
plan_output = planner.invoke({"initial_user_message": state["messages"]})
return {
"plan": plan_output.steps,
"messages": []
}
def prepare_next_step(state: TaskState):
"""Prepare the state for the executor LLM call for the next plan step."""
plan = state["plan"]
original_question = state["question"]
current_task = plan[0] if plan else ""
remaining_plan = plan[1:] if plan else []
task_message_content = f"""Original User Question: {original_question}
Current Task: {current_task}
Based *only* on the 'Current Task' description above, decide if a tool needs to be called.
If you call an analysis tool, pass the necessary arguments.
If no tool is needed for the Current Task, explain the reasoning or result based on the task description.
"""
task_message = HumanMessage(content=task_message_content)
updated_messages = state.get("messages", []) + [task_message]
return {
"plan": remaining_plan,
"current_task": current_task,
"messages": updated_messages
}
def executor_llm_call(state: TaskState):
"""Invoke the LLM with the current task, deciding on tool use."""
model_with_tools = llm.bind_tools(tools)
response = model_with_tools.invoke(state["messages"])
return {"messages": [response]}
def replan_step(state: TaskState):
"""Replans based on the completed step's result and history."""
current_task = state["current_task"]
messages = state["messages"]
latest_result = ""
if messages:
last_message = messages[-1]
if isinstance(last_message, AIMessage):
latest_result = last_message.content
elif isinstance(last_message, ToolMessage):
latest_result = last_message.content
else:
latest_result = str(last_message)
else:
latest_result = "(No message found for task result)"
past_steps_str = "\n".join(
f"Step: {task}\nThought: {thought}" for task, thought in state.get("past_steps", [])
)
plan_str = "\n".join(f"{i+1}. {step}" for i, step in enumerate(state.get("plan", [])))
replanner_input = {
"question": state["question"],
"plan_str": plan_str,
"past_steps_str": past_steps_str,
"current_task": current_task,
"latest_result": latest_result,
}
output = replanner.invoke(replanner_input)
updated_past_steps = [(current_task, output.thought)]
if isinstance(output.action, Response):
print(f"Replanner provided a final response: {output.action.response}")
final_answer_prompt = f"""The user's original question was: {state['question']}
The result determined by the plan is: {output.action.response}
Based on this result, output ONLY the final formatted answer itself, and nothing else.
Keep the answer concise and exact."""
final_answer_llm = get_llm()
extracted_response = final_answer_llm.invoke(final_answer_prompt).content.strip()
return {
"response": extracted_response,
"past_steps": updated_past_steps,
"messages": [],
"current_task": ""
}
else:
return {
"plan": output.action.steps,
"past_steps": updated_past_steps,
"messages": state["messages"],
"current_task": ""
}
# --- Conditional Routing Functions ---
def route_after_executor_call(state: TaskState) -> Literal["tool_node", "replan_step"]:
"""Route to tool node if tool call exists, otherwise to replan."""
messages = state["messages"]
last_message = messages[-1] if messages else None
if isinstance(last_message, AIMessage) and last_message.tool_calls:
return "tool_node"
else:
return "replan_step"
def route_after_replan(state: TaskState) -> Literal["prepare_next_step", END]:
"""Route to prepare next step if plan exists, otherwise end."""
if state.get("response"):
return END
elif state.get("plan"):
return "prepare_next_step"
else:
print("Warning: Replanner finished without response or new plan.")
return END
# --- File Handling Functions ---
def download_file(task_id: str, file_name: str, api_url: str = DEFAULT_API_URL) -> str:
"""Downloads file, returns path or empty string on failure."""
temp_dir = get_task_temp_dir(task_id)
file_url = f"{api_url}/files/{task_id}"
file_path = os.path.join(temp_dir, file_name)
try:
response = requests.get(file_url, stream=True)
response.raise_for_status()
with open(file_path, 'wb') as f:
for chunk in response.iter_content(chunk_size=8192):
f.write(chunk)
print(f"File downloaded successfully to {file_path}")
return file_path
except Exception as e:
print(f"Error downloading file: {str(e)}")
return ""
def process_file(state: TaskState):
"""Download file if needed, prepare initial state and message."""
task_id = state.get("task_id", "")
file_name = state.get("file_name", "")
api_url = state.get("api_url", DEFAULT_API_URL)
question = state.get("question", "")
initial_message_content = question
file_path_update = {}
temp_dir_update = {}
if task_id and file_name:
temp_dir = get_task_temp_dir(task_id)
temp_dir_update = {"temp_dir": temp_dir}
file_path = download_file(task_id, file_name, api_url)
file_path_update = {"file_path": file_path}
if file_path:
initial_message_content += f"\n\n(Note: File downloaded to: {file_path})"
else:
initial_message_content += f"\n\n(Note: Failed to download file '{file_name}')"
return {
"question": question,
"task_id": task_id,
"file_name": file_name,
"api_url": api_url,
**file_path_update,
**temp_dir_update,
"messages": [HumanMessage(content=initial_message_content)],
"plan": [],
"past_steps": [],
"response": "",
"current_task": "",
}
def process_input(state: TaskState) -> TaskState:
"""Prepare initial state when no file processing is needed."""
question = state.get("question", "")
return {
"question": question,
"task_id": state.get("task_id", ""),
"file_name": None,
"api_url": state.get("api_url", DEFAULT_API_URL),
"file_path": None,
"temp_dir": None,
"messages": [HumanMessage(content=question)],
"plan": [],
"past_steps": [],
"response": "",
"current_task": "",
}
def should_process_file(state: TaskState) -> Literal["process_file", "process_input"]:
"""Determine entry point based on file presence."""
task_id = state.get("task_id", "")
file_name = state.get("file_name", "")
if task_id and file_name:
return "process_file"
return "process_input"
# --- Build Graph ---
def create_plan_execute_task_flow():
"""Creates the LangGraph StateGraph for plan-and-execute agent."""
graph = StateGraph(TaskState)
# Add nodes
graph.add_node("process_input", process_input)
graph.add_node("process_file", process_file)
graph.add_node("planner", plan_step)
graph.add_node("prepare_next_step", prepare_next_step)
graph.add_node("executor_llm_call", executor_llm_call)
graph.add_node("tool_node", tool_node)
graph.add_node("replan_step", replan_step)
# Define edges
graph.set_conditional_entry_point(
should_process_file,
{"process_file": "process_file", "process_input": "process_input"}
)
graph.add_edge("process_input", "planner")
graph.add_edge("process_file", "planner")
graph.add_edge("planner", "prepare_next_step")
graph.add_edge("prepare_next_step", "executor_llm_call")
graph.add_conditional_edges(
"executor_llm_call",
route_after_executor_call,
{"tool_node": "tool_node", "replan_step": "replan_step"}
)
graph.add_edge("tool_node", "replan_step")
graph.add_conditional_edges(
"replan_step",
route_after_replan,
{"prepare_next_step": "prepare_next_step", END: END}
)
app = graph.compile()
print("Plan-and-execute task graph compiled.")
return app, graph
# --- LangGraph Agent Wrapper ---
class LangGraphAgent:
def __init__(self):
print("LangGraphAgent initialized with Plan-and-Execute flow.")
self.app_executor, _ = create_plan_execute_task_flow()
def __call__(self, item: dict) -> str:
task_id = item.get("task_id")
question = item.get("question")
file_name = item.get("file_name", None)
print(f"Agent received task {task_id}: {question[:50]}... (File: {file_name})")
if not question:
return "Error: Missing question in task item."
try:
initial_state = {
"task_id": task_id,
"question": question,
"file_name": file_name if file_name else None,
"api_url": DEFAULT_API_URL
}
print(f"Invoking agent for task {task_id}...")
result = self.app_executor.invoke(initial_state)
answer = result.get("response", "Error: No final response generated.")
if not isinstance(answer, str):
answer = str(answer)
print(f"Agent returning answer for task {task_id}: {answer[:50]}...")
return answer
except Exception as e:
print(f"Error processing task {task_id}: {e}")
import traceback
traceback.print_exc()
return f"Error: {str(e)}"
# --- Gradio Interface Functions ---
def run_and_submit_all(profile: gr.OAuthProfile | None):
"""Fetches all questions, runs the agent, submits all answers."""
space_id = os.getenv("SPACE_ID")
if not profile:
return "Please Login to Hugging Face with the button.", None
username = profile.username
print(f"User logged in: {username}")
api_url = DEFAULT_API_URL
questions_url = f"{api_url}/questions"
submit_url = f"{api_url}/submit"
try:
agent = LangGraphAgent()
except Exception as e:
print(f"Error instantiating agent: {e}")
return f"Error initializing agent: {e}", None
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
# Fetch questions
try:
response = requests.get(questions_url, timeout=15)
response.raise_for_status()
questions_data = response.json()
if not questions_data:
return "Fetched questions list is empty.", None
print(f"Fetched {len(questions_data)} questions.")
except Exception as e:
return f"Error fetching questions: {e}", None
# Run agent on questions
results_log = []
answers_payload = []
print(f"Running agent on {len(questions_data)} questions...")
for item in questions_data:
task_id = item.get("task_id")
question_text = item.get("question")
if not task_id or question_text is None:
continue
try:
submitted_answer = agent(item)
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
except Exception as e:
print(f"Error on task {task_id}: {e}")
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"ERROR: {e}"})
if not answers_payload:
return "No answers produced.", pd.DataFrame(results_log)
# Submit answers
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
try:
response = requests.post(submit_url, json=submission_data, timeout=120)
response.raise_for_status()
result_data = response.json()
final_status = (
f"✅ Submission Successful!\n"
f"User: {result_data.get('username')}\n"
f"Score: {result_data.get('score', 'N/A')}% "
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)"
)
return final_status, pd.DataFrame(results_log)
except Exception as e:
return f"Submission failed: {e}", pd.DataFrame(results_log)
# --- Gradio Interface ---
with gr.Blocks() as demo:
gr.Markdown("# 🦾 GAIA Agent Evaluator - Gemini Edition")
gr.Markdown(
"""
**Instructions:**
1. Login to Hugging Face
2. Click 'Run Evaluation & Submit'
3. Wait for the agent to process all questions
**Model:** Gemini 2.0 Flash Thinking (gratuit, excellent pour le raisonnement)
"""
)
gr.LoginButton()
run_button = gr.Button("🚀 Run Evaluation & Submit", variant="primary")
status_output = gr.Textbox(label="Status", lines=5, interactive=False)
results_table = gr.DataFrame(label="Results", wrap=True)
run_button.click(
fn=run_and_submit_all,
outputs=[status_output, results_table]
)
if __name__ == "__main__":
demo.launch()