agentCourse / app.py
gabejavitt's picture
Update app.py
0012c6a verified
import os
import io
import subprocess
import json
import re
import traceback
import contextlib
import uuid
import time
import ast
from typing import List, Optional, TypedDict, Annotated, Dict
from pathlib import Path
from collections import Counter
import gradio as gr
import pandas as pd
import numpy as np
import torch
from pydantic import BaseModel, Field
# Multimodal & Web Tools
from transformers import pipeline
from youtube_transcript_api import YouTubeTranscriptApi
from bs4 import BeautifulSoup
import requests
from PIL import Image
import base64
from googleapiclient.discovery import build
from googleapiclient.errors import HttpError
import assemblyai as aai
# LangChain & LangGraph
from langgraph.graph.message import add_messages
from langchain_core.messages import HumanMessage, AIMessage, ToolMessage, SystemMessage, AnyMessage, ToolCall
from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode
from langgraph.graph import START, END, StateGraph
from langchain_groq import ChatGroq
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_community.llms import HuggingFaceHub
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
# RAG
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.tools import DuckDuckGoSearchRun
from langchain_core.documents import Document
# =============================================================================
# CONFIGURATION
# =============================================================================
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
MAX_TURNS = 25
MAX_MESSAGE_LENGTH = 8000
REFLECT_EVERY_N_TURNS = 5
# =============================================================================
# GLOBAL RAG COMPONENTS
# =============================================================================
global_embeddings = None
global_text_splitter = None
def initialize_rag_components():
"""Initialize RAG components globally."""
global global_embeddings, global_text_splitter
if global_embeddings is None:
print("Initializing RAG embeddings...")
try:
global_embeddings = HuggingFaceEmbeddings(
model_name="sentence-transformers/all-MiniLM-L6-v2",
model_kwargs={'device': 'cpu'}
)
print("✅ Embeddings initialized.")
except Exception as e:
print(f"⚠️ Failed to initialize embeddings: {e}")
return False
if global_text_splitter is None:
print("Initializing text splitter...")
global_text_splitter = RecursiveCharacterTextSplitter(
chunk_size=500,
chunk_overlap=50,
length_function=len,
separators=["\n\n", "\n", ". ", " ", ""]
)
print("✅ Text splitter initialized.")
return True
# =============================================================================
# ANSWER SHEET VALIDATION FUNCTIONS
# =============================================================================
def load_answer_sheet(filepath: str = "answer_sheet.json") -> Dict[str, str]:
"""Load the answer sheet from a JSON file"""
try:
if os.path.exists(filepath):
with open(filepath, 'r', encoding='utf-8') as f:
answers = json.load(f)
print(f"✅ Loaded answer sheet with {len(answers)} answers from {filepath}")
return answers
else:
print(f"⚠️ Answer sheet not found at {filepath}")
return {}
except Exception as e:
print(f"❌ Error loading answer sheet: {e}")
return {}
def check_answer_correctness(submitted: str, correct: str) -> tuple[bool, str]:
"""
Check if submitted answer matches correct answer with fuzzy matching
Returns: (is_correct, feedback_message)
"""
# Normalize both answers
submitted_norm = submitted.strip().lower()
correct_norm = correct.strip().lower()
# Exact match
if submitted_norm == correct_norm:
return True, "✅ EXACT MATCH"
# Remove common punctuation and check again
import string
submitted_clean = submitted_norm.translate(str.maketrans('', '', string.punctuation))
correct_clean = correct_norm.translate(str.maketrans('', '', string.punctuation))
if submitted_clean == correct_clean:
return True, "✅ MATCH (punctuation difference)"
# Check if it's a number formatting issue
try:
# Try to parse as numbers
submitted_num = float(submitted_clean.replace(',', '').replace('$', ''))
correct_num = float(correct_clean.replace(',', '').replace('$', ''))
if abs(submitted_num - correct_num) < 0.01: # Allow small floating point differences
return True, "✅ MATCH (numeric equivalence)"
except (ValueError, AttributeError):
pass
# Check if submitted answer contains correct answer (for list-type answers)
if ',' in correct_norm:
correct_items = set([item.strip() for item in correct_norm.split(',')])
submitted_items = set([item.strip() for item in submitted_norm.split(',')])
if correct_items == submitted_items:
return True, "✅ MATCH (item order difference)"
missing_items = correct_items - submitted_items
extra_items = submitted_items - correct_items
if missing_items and not extra_items:
return False, f"❌ MISSING: {', '.join(missing_items)}"
elif extra_items and not missing_items:
return False, f"❌ EXTRA: {', '.join(extra_items)}"
elif missing_items and extra_items:
return False, f"❌ MISSING: {', '.join(missing_items)} | EXTRA: {', '.join(extra_items)}"
# Check case-insensitive substring match
if submitted_norm in correct_norm or correct_norm in submitted_norm:
return False, f"❌ PARTIAL MATCH (submitted: '{submitted}' | correct: '{correct}')"
return False, f"❌ WRONG (submitted: '{submitted}' | correct: '{correct}')"
def create_answer_sheet_template(questions: List[Dict], filepath: str = "answer_sheet.json"):
"""Create an answer sheet template from questions"""
answer_template = {}
for q in questions:
answer_template[q['task_id']] = ""
with open(filepath, 'w', encoding='utf-8') as f:
json.dump(answer_template, f, indent=2)
print(f"✅ Created answer sheet template at {filepath}")
print(f" Please fill in the correct answers for {len(answer_template)} questions")
# =============================================================================
# ASR INITIALIZATION
# =============================================================================
asr_pipeline = None
try:
print("Loading ASR (Whisper) pipeline globally...")
device = 0 if torch.cuda.is_available() else -1
device_name = "cuda:0" if device == 0 else "cpu"
print(f"Attempting to use device: {device_name} for ASR.")
asr_pipeline = pipeline(
"automatic-speech-recognition",
model="openai/whisper-base",
torch_dtype=torch.float16 if device == 0 else torch.float32,
device=device
)
print("✅ ASR (Whisper) pipeline loaded successfully.")
except Exception as e:
print(f"⚠️ Warning: Could not load ASR pipeline globally. Error: {e}")
asr_pipeline = None
# =============================================================================
# UTILITY FUNCTIONS
# =============================================================================
def remove_fences_simple(text):
"""Remove code fences from text."""
original_text = text
text = text.strip()
if text.startswith("```") and text.endswith("```"):
text = text[3:-3].strip()
if '\n' in text:
first_line, rest = text.split('\n', 1)
if first_line.strip().replace('_','').isalnum() and len(first_line.strip()) < 15:
text = rest.strip()
return text
return original_text
def truncate_if_needed(content: str, max_length: int = MAX_MESSAGE_LENGTH) -> str:
"""Truncate content if it exceeds max length."""
if len(content) > max_length:
return content[:max_length] + f"\n...[truncated, {len(content)} total chars]"
return content
def find_file(path: str) -> Optional[Path]:
"""Find a file by trying multiple path variations."""
script_dir = Path.cwd()
safe_path = Path(path).as_posix()
paths_to_try = [
script_dir / safe_path,
Path(safe_path),
script_dir / Path(path).name
]
for attempt_path in paths_to_try:
if attempt_path.exists():
return attempt_path
return None
# =============================================================================
# PLANNING & REFLECTION TOOLS
# =============================================================================
class ThinkInput(BaseModel):
reasoning: str = Field(description="Brief reasoning summary (under 150 chars)")
@tool(args_schema=ThinkInput)
def think_through_logic(reasoning: str) -> str:
"""
Use this to work through logic puzzles, riddles, or reasoning problems.
Call this when:
- The question is a riddle or brain teaser
- You need to reason through a logical problem
- No external information is needed, just thinking
After thinking, use calculator if math is involved, then validate and submit answer.
"""
print(f"🧠 Thinking: {reasoning[:100]}...")
return f"""✅ Logic reasoning recorded.
Next steps:
1. If math needed → use calculator()
2. Once you have answer → use validate_answer()
3. Then → use final_answer_tool()
Remember: You MUST call another tool. Do not output reasoning text."""
class PlanInput(BaseModel):
task_summary: str = Field(description="Very brief task summary (under 80 chars)")
@tool(args_schema=PlanInput)
def create_plan(task_summary: str) -> str:
"""
Creates a plan for multi-step questions. Use for complex tasks only.
Keep the summary VERY brief to avoid errors.
"""
print(f"📋 Planning: {task_summary[:80]}...")
return f"""✅ Plan created for: {task_summary}
FRAMEWORK:
1. What info do I need?
2. What tools will I use?
3. In what order?
Now execute step 1. You MUST call a tool next."""
class ReflectInput(BaseModel):
situation: str = Field(description="Brief situation summary (under 80 chars)")
@tool(args_schema=ReflectInput)
def reflect_on_progress(situation: str) -> str:
"""
Reflects on progress when stuck. Use after 5+ turns without progress.
Keep situation summary VERY brief.
"""
print(f"🤔 Reflecting: {situation[:80]}...")
return f"""🔍 REFLECTION on: {situation}
QUESTIONS:
1. Am I using the right approach?
2. Should I try a different tool?
3. Do I actually have the answer already?
Take a DIFFERENT approach now. You MUST call a tool next."""
class ValidateInput(BaseModel):
proposed_answer: str = Field(description="The answer to validate")
original_question: str = Field(description="Original question (first 100 chars)")
@tool(args_schema=ValidateInput)
def validate_answer(proposed_answer: str, original_question: str) -> str:
"""
Validates answer format before submission. ALWAYS use before final_answer_tool.
"""
print(f"✓ Validating: '{proposed_answer[:50]}...'")
issues = []
warnings = []
# Check for conversational fluff
fluff = ["the answer is", "based on", "according to", "i found", "here is"]
if any(p in proposed_answer.lower() for p in fluff):
issues.append("❌ Remove conversational text. Answer only.")
# Check for code fences
if "```" in proposed_answer:
issues.append("❌ Remove code fences (```).")
# Check length
if len(proposed_answer) > 500:
warnings.append("⚠️ Answer very long. Just the answer?")
# Check for number questions
if any(k in original_question.lower() for k in ["how many", "what number", "count"]):
if not any(c.isdigit() for c in proposed_answer):
warnings.append("⚠️ Question asks for number but answer has no digits.")
if issues:
return "🚫 VALIDATION FAILED:\n" + "\n".join(issues) + "\n\nFix then retry."
if warnings:
return "⚠️ WARNINGS:\n" + "\n".join(warnings) + "\n\nConsider fixing, or proceed if confident."
return "✅ VALIDATION PASSED! Now call final_answer_tool() with this answer."
# =============================================================================
# CORE TOOLS
# =============================================================================
class SearchInput(BaseModel):
query: str = Field(description="Search query (concise)")
@tool(args_schema=SearchInput)
def search_tool(query: str) -> str:
"""
Search the web for information. Returns snippets.
IMPORTANT: Search results are SNIPPETS only. For complete information:
1. Use search_tool to find URLs
2. Use scrape_and_retrieve to get FULL page content
Example workflow:
- search_tool("Mercedes Sosa Wikipedia") → get URL
- scrape_and_retrieve(url=..., query="studio albums 2000-2009")
"""
if not isinstance(query, str) or not query.strip():
return "Error: Invalid query."
# Auto-add Wikipedia site filter if mentioned
if 'wikipedia' in query.lower() and 'site:' not in query:
query = f"{query} site:wikipedia.org"
print(f"🔍 Searching: {query}")
max_retries = 3
for attempt in range(max_retries):
try:
search = DuckDuckGoSearchRun()
result = search.run(query)
if not result or len(result) < 50:
return "No relevant results found. Try different search terms or check if the information exists."
return truncate_if_needed(result)
except Exception as e:
if attempt < max_retries - 1:
time.sleep(2 ** attempt)
continue
return f"Search error after {max_retries} attempts: {str(e)}"
class CalcInput(BaseModel):
expression: str = Field(description="Math expression (e.g., '2+2', 'sqrt(16)')")
@tool(args_schema=CalcInput)
def calculator(expression: str) -> str:
"""
Evaluates math expressions. Use for ANY calculations.
Supports: +, -, *, /, **, sqrt, sin, cos, log, pi, e, etc.
"""
if not isinstance(expression, str) or not expression.strip():
return "Error: Invalid expression."
print(f"🧮 Calculating: {expression}")
try:
import math
safe_dict = {
'sqrt': math.sqrt, 'sin': math.sin, 'cos': math.cos, 'tan': math.tan,
'log': math.log, 'log10': math.log10, 'exp': math.exp,
'pi': math.pi, 'e': math.e, 'abs': abs, 'round': round,
'pow': pow, 'sum': sum, 'min': min, 'max': max
}
result = eval(expression, {"__builtins__": {}}, safe_dict)
return str(result)
except Exception as e:
return f"Calculation error for '{expression}': {str(e)}"
class CodeInput(BaseModel):
code: str = Field(description="Python code (MUST include print() for output)")
@tool(args_schema=CodeInput)
def code_interpreter(code: str) -> str:
"""
Executes Python code with timeout protection.
CRITICAL: Always use print() to output results!
"""
if not isinstance(code, str):
return "Error: code must be string."
# Safety checks
dangerous = ['__import__', 'eval(', 'compile(', 'subprocess', 'os.system', 'exec(']
if any(d in code.lower() for d in dangerous):
return f"Error: Dangerous operation not allowed."
if 'open(' in code.lower() and any(m in code for m in ["'w'", '"w"', "'a'", '"a"']):
return "Error: File writing not allowed. Use write_file tool."
print(f"💻 Executing code ({len(code)} chars)...")
output_stream = io.StringIO()
error_stream = io.StringIO()
try:
with contextlib.redirect_stdout(output_stream), contextlib.redirect_stderr(error_stream):
safe_globals = {
"pd": pd,
"np": np,
"json": json,
"re": re,
"__builtins__": __builtins__
}
exec(code, safe_globals, {})
stdout = output_stream.getvalue()
stderr = error_stream.getvalue()
if stderr:
return f"Error:\n{stderr}\n\nStdout:\n{stdout}"
if stdout:
return truncate_if_needed(stdout)
return "Code executed but no output. Remember to use print()!"
except Exception as e:
return f"Execution failed:\n{traceback.format_exc()}"
class ReadFileInput(BaseModel):
path: str = Field(description="File path")
@tool(args_schema=ReadFileInput)
def read_file(path: str) -> str:
"""Reads file content."""
if not isinstance(path, str) or not path.strip():
return "Error: Invalid path."
print(f"📄 Reading: {path}")
file_path = find_file(path)
if not file_path:
return f"Error: File not found: '{path}'\nCWD files: {os.listdir('.')}"
try:
content = file_path.read_text(encoding='utf-8')
return truncate_if_needed(content)
except UnicodeDecodeError:
return f"Error: Binary file. Size: {file_path.stat().st_size} bytes. Try audio_transcription_tool for audio."
except Exception as e:
return f"Read error: {str(e)}"
class WriteFileInput(BaseModel):
path: str = Field(description="File path")
content: str = Field(description="Content to write")
@tool(args_schema=WriteFileInput)
def write_file(path: str, content: str) -> str:
"""Writes content to file."""
if not path or not isinstance(content, str):
return "Error: Invalid inputs."
print(f"✍️ Writing: {path}")
try:
file_path = Path.cwd() / path
file_path.parent.mkdir(parents=True, exist_ok=True)
file_path.write_text(content, encoding='utf-8')
return f"Wrote {len(content)} chars to '{path}'."
except Exception as e:
return f"Write error: {str(e)}"
class ListDirInput(BaseModel):
path: str = Field(description="Directory path", default=".")
@tool(args_schema=ListDirInput)
def list_directory(path: str = ".") -> str:
"""Lists directory contents."""
print(f"📁 Listing: {path}")
try:
dir_path = Path.cwd() / path if path != "." else Path.cwd()
if not dir_path.is_dir():
return f"Error: '{path}' not a directory."
items = sorted(dir_path.iterdir())
if not items:
return f"Directory '{path}' is empty."
files, dirs = [], []
for item in items:
if item.is_dir():
dirs.append(f"📁 {item.name}/")
else:
files.append(f"📄 {item.name} ({item.stat().st_size} bytes)")
result = f"Contents of '{path}':\n\n"
if dirs:
result += "Directories:\n" + "\n".join(dirs) + "\n\n"
if files:
result += "Files:\n" + "\n".join(files)
return result
except Exception as e:
return f"List error: {str(e)}"
class AudioInput(BaseModel):
file_path: str = Field(description="Audio file path")
@tool(args_schema=AudioInput)
def audio_transcription_tool(file_path: str) -> str:
"""Transcribes audio using Whisper."""
if not file_path:
return "Error: Invalid file path."
print(f"🎤 Transcribing: {file_path}")
if asr_pipeline is None:
return "Error: ASR not available."
audio_path = find_file(file_path)
if not audio_path:
return f"Error: Audio file not found: '{file_path}'"
try:
transcription = asr_pipeline(
str(audio_path),
return_timestamps=True, # ← Add this!
chunk_length_s=30, # ← Process in 30-second chunks
stride_length_s=5 # ← 5-second overlap between chunks
)
# Extract just the text (ignore timestamps)
result_text = transcription.get("text", "")
# OR if you want to see the chunks:
# chunks = transcription.get("chunks", [])
# result_text = " ".join([chunk["text"] for chunk in chunks])
if not result_text:
return "Error: Transcription empty."
return f"Transcription:\n{truncate_if_needed(result_text)}"
except Exception as e:
return f"Transcription error: {str(e)}"
class ImageAnalysisInput(BaseModel):
file_path: str = Field(description="Image file path")
query: str = Field(description="What to analyze in the image")
@tool(args_schema=ImageAnalysisInput)
def analyze_image(file_path: str, query: str) -> str:
"""
Analyzes images using Google Gemini Vision API.
Use for: chess positions, diagrams, charts, photos, screenshots.
Provide the EXACT file path from [FILE ATTACHED: ...] in the question.
"""
if not file_path or not query:
return "Error: file_path and query required."
print(f"🖼️ Analyzing image: {file_path}")
print(f" Query: {query[:100]}...")
# Try to find the file
image_path = find_file(file_path)
# If not found via find_file, try the path directly (for /tmp files)
if not image_path and os.path.exists(file_path):
image_path = Path(file_path)
if not image_path or not image_path.exists():
return f"Error: Image not found at '{file_path}'. Check [FILE ATTACHED: ...] in question for correct path."
print(f"✓ Found image at: {image_path}")
try:
GOOGLE_API_KEY = os.getenv("GEMINI_API_KEY")
if not GOOGLE_API_KEY:
return "Error: GEMINI_API_KEY not set."
# Load and encode image
img = Image.open(image_path)
print(f" Image size: {img.size}, mode: {img.mode}")
# Convert to RGB if necessary
if img.mode not in ['RGB', 'RGBA']:
img = img.convert('RGB')
# Convert to base64
buffered = io.BytesIO()
img.save(buffered, format="JPEG")
img_base64 = base64.b64encode(buffered.getvalue()).decode()
print(f" Encoded image: {len(img_base64)} bytes")
# Use Gemini Vision
vision_llm = ChatGoogleGenerativeAI(
model="gemini-2.0-flash",
google_api_key=GOOGLE_API_KEY,
temperature=0
)
message = HumanMessage(
content=[
{"type": "text", "text": query},
{
"type": "image_url",
"image_url": f"data:image/jpeg;base64,{img_base64}"
}
]
)
print(f" Sending to Gemini Vision...")
response = vision_llm.invoke([message])
print(f"✓ Got response: {len(response.content)} chars")
return f"Image Analysis:\n{truncate_if_needed(response.content)}"
except Exception as e:
error_msg = f"Image analysis error: {str(e)}"
print(f"❌ {error_msg}")
print(traceback.format_exc())
return error_msg
class YoutubeInput(BaseModel):
video_url: str = Field(description="YouTube URL")
@tool(args_schema=YoutubeInput)
def get_youtube_transcript(video_url: str) -> str:
"""
Fetches YouTube video transcript using AssemblyAI.
Works reliably on Hugging Face Spaces.
"""
try:
# Set API key (store in HF Spaces secrets)
aai.settings.api_key = os.getenv("ASSEMBLYAI_API_KEY")
print(f"📺 Transcribing: {video_url}")
# Transcribe directly from YouTube URL
transcriber = aai.Transcriber()
transcript = transcriber.transcribe(video_url)
# Wait for transcription
if transcript.status == aai.TranscriptStatus.error:
return f"Error: {transcript.error}"
print(f"✓ Transcribed {len(transcript.text)} chars")
return f"Transcript:\n{transcript.text}"
except Exception as e:
return f"Error: {str(e)}"
class ScrapeInput(BaseModel):
url: str = Field(description="URL (must start with http:// or https://)")
query: str = Field(description="Specific information to find on the page")
@tool(args_schema=ScrapeInput)
def scrape_and_retrieve(url: str, query: str) -> str:
"""
Fetch and search FULL webpage content using RAG (not just snippets like search_tool).
CRITICAL: Use this after search_tool gives you a URL. This gets the COMPLETE page.
Workflow Example:
1. search_tool('Mercedes Sosa Wikipedia') → get URL
2. scrape_and_retrieve(
url='https://en.wikipedia.org/wiki/Mercedes_Sosa',
query='studio albums released between 2000 and 2009'
) → Returns FULL discography section
Use when:
- Counting items (albums, people, events, etc.)
- Finding specific names, dates, or numbers
- Need complete tables or lists
- Wikipedia articles, documentation, papers
- Search snippets weren't enough
"""
if not url.startswith(('http://', 'https://')):
return f"Error: Invalid URL format. Must start with http:// or https://"
if not query:
return "Error: Query required to search the page content."
if global_embeddings is None or global_text_splitter is None:
if not initialize_rag_components():
return "Error: RAG components not initialized."
print(f"🌐 Scraping: {url}")
print(f" Looking for: {query[:100]}...")
max_retries = 3
for attempt in range(max_retries):
try:
headers = {
'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'
}
response = requests.get(url, headers=headers, timeout=20)
response.raise_for_status()
soup = BeautifulSoup(response.text, 'html.parser')
# Remove noise
for tag in soup(["script", "style", "nav", "footer", "aside", "header", "iframe"]):
tag.extract()
# Extract main content
main = soup.find('main') or soup.find('article') or soup.find('div', class_='mw-parser-output') or soup.body
if not main:
return "Error: Could not find main content on page."
text = main.get_text(separator='\n', strip=True)
lines = [l.strip() for l in text.splitlines() if l.strip()]
text = '\n'.join(lines)
if len(text) < 50:
return f"Error: Page content too short ({len(text)} chars). May be blocked or empty."
print(f"✓ Extracted {len(text)} characters from page")
# Chunk and search
chunks = global_text_splitter.split_text(text)
if not chunks:
return "Error: Could not process page content."
print(f"✓ Created {len(chunks)} chunks")
docs = [Document(page_content=c, metadata={"source": url}) for c in chunks]
db = FAISS.from_documents(docs, global_embeddings)
retriever = db.as_retriever(search_kwargs={"k": 5})
retrieved = retriever.invoke(query)
if not retrieved:
return f"No information found matching: '{query}'\nTry a different query or the information may not be on this page."
print(f"✓ Found {len(retrieved)} relevant chunks")
context = "\n\n---\n\n".join([f"[Section {i+1}]\n{d.page_content}" for i, d in enumerate(retrieved)])
return truncate_if_needed(f"From {url}:\n\n{context}")
except requests.Timeout:
if attempt < max_retries - 1:
print(f"⚠️ Timeout, retrying... (attempt {attempt + 1}/{max_retries})")
time.sleep(2 ** attempt)
continue
return f"Error: Page request timed out after {max_retries} attempts."
except requests.RequestException as e:
if attempt < max_retries - 1:
time.sleep(2 ** attempt)
continue
return f"Error fetching page: {str(e)}"
except Exception as e:
return f"Error processing page: {str(e)}\n{traceback.format_exc()}"
class ChessAnalysisInput(BaseModel):
image_path: str = Field(description="Path to chess board image file")
description: str = Field(description="Any additional context about the position (optional)", default="")
@tool(args_schema=ChessAnalysisInput)
@tool(args_schema=ChessAnalysisInput)
def analyze_chess_position(image_path: str, description: str = "") -> str:
"""
Analyzes a chess position from an image using Stockfish engine.
MUCH MORE RELIABLE than Lichess API because:
- Works offline
- Analyzes ANY position (not just cloud database)
- Stronger engine (Stockfish 16+)
- No rate limits or 404 errors
Use this tool when:
- Question mentions chess, checkmate, or chess notation
- An image file shows a chess board
- Need to find the best move in a position
Args:
image_path: Path to chess board image
description: The full question text - IMPORTANT for determining whose turn it is!
Returns: Best move in algebraic notation (e.g., "Qh5", "Nf6+", "Rd5")
"""
if not image_path:
return "Error: image_path is required."
print(f"♟️ Analyzing chess position from: {image_path}")
# Find the file
chess_image = find_file(image_path)
# If not found via find_file, try direct path
if not chess_image and os.path.exists(image_path):
chess_image = Path(image_path)
if not chess_image or not chess_image.exists():
return f"Error: Chess board image not found at '{image_path}'. Check the [FILE ATTACHED: ...] path in the question."
print(f"✓ Found chess image at: {chess_image}")
try:
# ====================================================================
# STEP 1: Extract FEN notation from image using Gemini Vision
# ====================================================================
GOOGLE_API_KEY = os.getenv("GEMINI_API_KEY")
if not GOOGLE_API_KEY:
return "Error: GEMINI_API_KEY not set in Space secrets."
print("📸 Extracting chess position from image using Gemini...")
# Load and encode image
img = Image.open(chess_image)
print(f" Image loaded: {img.size}, mode: {img.mode}")
if img.mode not in ['RGB', 'RGBA']:
img = img.convert('RGB')
buffered = io.BytesIO()
img.save(buffered, format="JPEG")
img_base64 = base64.b64encode(buffered.getvalue()).decode()
# Use Gemini Vision to extract FEN
vision_llm = ChatGoogleGenerativeAI(
model="gemini-2.5-pro",
google_api_key=GOOGLE_API_KEY,
temperature=0
)
# Check if the question explicitly states whose turn it is
whose_turn = None
if description:
desc_lower = description.lower()
if "black" in desc_lower and ("turn" in desc_lower or "move" in desc_lower):
whose_turn = "b"
elif "white" in desc_lower and ("turn" in desc_lower or "move" in desc_lower):
whose_turn = "w"
fen_prompt = f"""Analyze this chess board image and provide the position in FEN notation.
CRITICAL INSTRUCTIONS:
1. Carefully identify each piece:
- White pieces (UPPERCASE): K=King, Q=Queen, R=Rook, B=Bishop, N=Knight, P=Pawn
- Black pieces (lowercase): k, q, r, b, n, p
2. BOARD ORIENTATION - This is CRITICAL:
- In chess diagrams, the board is shown from the perspective of the player to move
- Look at the BOTTOM rank (closest to viewer):
* If bottom pieces are BLACK (lowercase in FEN) → Black to move → active color = 'b'
* If bottom pieces are WHITE (uppercase in FEN) → White to move → active color = 'w'
- The rank labels (1-8) on the side can help:
* If rank 8 is at bottom and rank 1 at top → Black's perspective → use 'b'
* If rank 1 is at bottom and rank 8 at top → White's perspective → use 'w'
{"- OVERRIDE: The question explicitly states BLACK's turn, so use 'b'" if whose_turn == "b" else ""}
{"- OVERRIDE: The question explicitly states WHITE's turn, so use 'w'" if whose_turn == "w" else ""}
3. FEN Format (read from rank 8 to rank 1, left to right):
- Use numbers (1-8) for consecutive empty squares
- Use '/' to separate ranks
- IMPORTANT: Always write FEN from White's perspective (rank 8 first, rank 1 last)
- But set the active_color based on whose perspective the board shows
4. Return ONLY the FEN string in this exact format:
piece_placement active_color castling en_passant halfmove fullmove
Example: rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
DOUBLE-CHECK:
- Did you identify whose pieces are at the BOTTOM of the board?
- Did you set active_color correctly based on board orientation?
- Did you write piece_placement from rank 8 to rank 1?
Return ONLY the FEN string, nothing else."""
message = HumanMessage(
content=[
{"type": "text", "text": fen_prompt},
{
"type": "image_url",
"image_url": f"data:image/jpeg;base64,{img_base64}"
}
]
)
response = vision_llm.invoke([message])
fen_raw = response.content.strip()
print(f"📝 Raw FEN response: {fen_raw}")
# Clean up FEN (remove markdown, explanations, etc.)
fen = None
for line in fen_raw.split('\n'):
line = line.strip().replace('```', '').replace('fen', '')
# FEN should have '/' for ranks and spaces for components
if '/' in line and ' ' in line and not line.startswith('#'):
if any(c in line for c in 'kqrbnpKQRBNP12345678'):
fen = line
break
if not fen:
return f"Error: Could not extract valid FEN notation from image. Response: {fen_raw[:200]}"
print(f"✓ Extracted FEN: {fen}")
# Override the turn indicator if we know from the question
if whose_turn:
fen_parts = fen.split()
if len(fen_parts) >= 2:
old_turn = fen_parts[1]
fen_parts[1] = whose_turn
fen = ' '.join(fen_parts)
print(f"🔄 Corrected turn from '{old_turn}' to '{whose_turn}' based on question")
print(f"✓ Corrected FEN: {fen}")
# Additional verification: Check if board orientation matches turn
# In FEN, rank 8 is first, rank 1 is last
# If bottom of image shows black pieces, it's black's turn
fen_parts = fen.split()
piece_placement = fen_parts[0]
active_color = fen_parts[1] if len(fen_parts) > 1 else 'w'
# Get last rank (rank 1 in FEN, which is bottom if white's perspective)
ranks = piece_placement.split('/')
rank_1 = ranks[-1] # Last rank in FEN
rank_8 = ranks[0] # First rank in FEN
# Check which color dominates bottom rank
# If showing from black's perspective, rank 8 should be at bottom
# and active color should be 'b'
black_pieces_in_rank8 = sum(1 for c in rank_8 if c.islower() and c.isalpha())
white_pieces_in_rank8 = sum(1 for c in rank_8 if c.isupper() and c.isalpha())
if black_pieces_in_rank8 > white_pieces_in_rank8 and active_color == 'w':
print(f"⚠️ Warning: Rank 8 has more black pieces, likely black's perspective")
print(f" Changing active color from 'w' to 'b'")
fen_parts[1] = 'b'
fen = ' '.join(fen_parts)
# ====================================================================
# STEP 2: Validate FEN with python-chess
# ====================================================================
try:
import chess
except ImportError:
return "Error: python-chess not installed. Add 'python-chess' to requirements.txt"
try:
board = chess.Board(fen)
print(f"✓ FEN validated successfully")
print(f" Turn: {'White' if board.turn else 'Black'}")
print(f" Legal moves: {board.legal_moves.count()}")
except ValueError as e:
return f"Error: Invalid FEN notation: {e}\nExtracted FEN: {fen}"
# ====================================================================
# STEP 3: Analyze with Stockfish
# ====================================================================
print("🔍 Analyzing position with Stockfish...")
try:
from stockfish import Stockfish
except ImportError:
return "Error: stockfish not installed. Add 'stockfish' to requirements.txt and install Stockfish binary"
# Try to find Stockfish binary
stockfish_paths = [
"/usr/games/stockfish", # Linux (apt-get install)
"/usr/local/bin/stockfish", # Mac (brew install)
"/usr/bin/stockfish", # Alternative Linux
"stockfish", # In PATH
"./stockfish", # Local directory
"C:\\Program Files\\stockfish\\stockfish.exe" # Windows
]
stockfish_path = None
for path in stockfish_paths:
if os.path.exists(path) or os.path.isfile(path):
stockfish_path = path
break
if not stockfish_path:
# Try running 'which stockfish' on Unix systems
try:
import subprocess
result = subprocess.run(['which', 'stockfish'],
capture_output=True,
text=True,
timeout=5)
if result.returncode == 0:
stockfish_path = result.stdout.strip()
except:
pass
if not stockfish_path:
return """Error: Stockfish binary not found. Install it:
- Linux: sudo apt-get install stockfish
- Mac: brew install stockfish
- Windows: Download from stockfishchess.org
Or set the path manually in the code."""
print(f"✓ Found Stockfish at: {stockfish_path}")
# Initialize Stockfish
try:
stockfish = Stockfish(
path=stockfish_path,
depth=35, # Analysis depth (higher = stronger but slower)
parameters={
"Threads": 2,
"Minimum Thinking Time": 5000, # milliseconds
"Hash": 1024, # MB of RAM
}
)
except Exception as e:
return f"Error initializing Stockfish: {e}"
# Set position
stockfish.set_fen_position(fen)
# Get best move
print(" Computing best move...")
best_move_uci = stockfish.get_best_move()
if not best_move_uci:
return "Error: Stockfish could not find a legal move. Check if position is valid."
print(f"🎯 Best move (UCI): {best_move_uci}")
# Get evaluation
evaluation = stockfish.get_evaluation()
eval_type = evaluation.get("type", "cp")
eval_value = evaluation.get("value", 0)
if eval_type == "mate":
eval_str = f" (Mate in {abs(eval_value)})"
else:
# Centipawns to pawns
eval_str = f" (Eval: {eval_value/100:+.2f})"
# ====================================================================
# STEP 4: Convert UCI to Standard Algebraic Notation (SAN)
# ====================================================================
try:
uci_move = chess.Move.from_uci(best_move_uci)
san_move = board.san(uci_move)
# Check if move leads to check/checkmate
board.push(uci_move)
if board.is_checkmate():
check_str = " - Checkmate!"
elif board.is_check():
check_str = " - Check"
else:
check_str = ""
final_result = f"{san_move}{eval_str}{check_str}"
print(f"✅ Best move: {final_result}")
# Return JUST the move notation for clean submission
return san_move
except Exception as e:
print(f"⚠️ Could not convert to SAN: {e}")
# Fall back to UCI notation
return best_move_uci
except Exception as e:
error_msg = f"Chess analysis failed: {str(e)}"
print(f"❌ {error_msg}")
print(traceback.format_exc())
return error_msg
class FinalAnswerInput(BaseModel):
answer: str = Field(description="Final answer - EXACTLY what was asked, nothing more")
@tool(args_schema=FinalAnswerInput)
def final_answer_tool(answer: str) -> str:
"""
Submit final answer. CRITICAL RULES:
1. ALWAYS call validate_answer() first
2. Answer must be EXACTLY what was asked
3. NO conversational text
4. NO explanations
5. Match requested format exactly
"""
if not isinstance(answer, str):
answer = str(answer)
print(f"✅ FINAL ANSWER SUBMITTED: {answer}")
return answer
# =============================================================================
# DEFINED TOOLS LIST
# =============================================================================
defined_tools = [
# Planning & Reflection
think_through_logic,
create_plan,
reflect_on_progress,
validate_answer,
# Core tools
search_tool,
calculator,
code_interpreter,
# File operations
read_file,
write_file,
list_directory,
# Specialized
audio_transcription_tool,
analyze_image,
get_youtube_transcript,
scrape_and_retrieve,
analyze_chess_position,
# Final
final_answer_tool
]
# =============================================================================
# AGENT STATE
# =============================================================================
class AgentState(TypedDict):
messages: Annotated[List[AnyMessage], add_messages]
turn: int
has_plan: bool
consecutive_errors: int
tool_history: List[str]
last_tool_was_thinking: bool
# =============================================================================
# ENHANCED FALLBACK PARSER
# =============================================================================
def parse_tool_call_from_string(content: str, tools: List) -> List[ToolCall]:
"""Enhanced parser with multiple strategies."""
print(f"🔧 Fallback parsing (first 300 chars):\n{content[:300]}")
tool_name = None
tool_input = None
# STRATEGY 1: Groq's <function=name{...}> format
groq_match = re.search(r"<function=(\w+)\s*(\{.*?\})\s*(?:>|</function>)", content, re.DOTALL)
if groq_match:
try:
tool_name = groq_match.group(1).strip()
json_str = groq_match.group(2).strip()
json_str = json_str.encode().decode('unicode_escape')
tool_input = json.loads(json_str)
print(f"✓ Parsed Groq format: {tool_name}")
except:
tool_name = None
# STRATEGY 2: Standard <function(name)>{...} format
if not tool_name:
func_match = re.search(r"<function[(=]\s*([^)]+)\s*[)>](.*)", content, re.DOTALL | re.IGNORECASE)
if func_match:
try:
tool_name = func_match.group(1).strip().replace("'", "").replace('"', '')
remaining = func_match.group(2)
json_start = remaining.find('{')
if json_start != -1:
json_str = remaining[json_start:].strip().rstrip(',')
tool_input = json.loads(json_str)
print(f"✓ Parsed standard format: {tool_name}")
except:
tool_name = None
# STRATEGY 3: Tool mention with code block → wrap in code_interpreter
if not tool_name and "```python" in content:
try:
code_match = re.search(r"```python\n(.*?)```", content, re.DOTALL)
if code_match:
code = code_match.group(1).strip()
tool_name = "code_interpreter"
tool_input = {"code": code}
print(f"✓ Extracted Python code → code_interpreter")
except:
pass
# STRATEGY 4: Direct tool mention → create minimal valid call
if not tool_name:
for tool in tools:
if tool.name.lower() in content.lower():
tool_name = tool.name
tool_input = {}
# Try to extract arguments from content
if tool.args_schema:
schema = tool.args_schema.model_json_schema()
for prop in schema.get('properties', {}).keys():
if prop in schema.get('required', []):
# Use placeholder
tool_input[prop] = "auto_extracted"
print(f"✓ Found mention of '{tool_name}' → creating default call")
break
# STRATEGY 5: Emergency - if no tool detected, force a reasonable one
if not tool_name:
# If content looks like reasoning, use think_through_logic
if len(content) > 50 and not any(kw in content.lower() for kw in ["error", "failed", "invalid"]):
tool_name = "think_through_logic"
tool_input = {"reasoning": content[:150]}
print(f"⚠️ No tool detected → forcing think_through_logic")
# Validate and create tool call
if tool_name and tool_input is not None:
matching_tools = [t for t in tools if t.name == tool_name]
if matching_tools:
return [ToolCall(name=tool_name, args=tool_input, id=str(uuid.uuid4()))]
else:
print(f"❌ Tool '{tool_name}' not in available tools")
print("❌ All parsing strategies failed")
return []
# =============================================================================
# CONDITIONAL EDGE FUNCTION
# =============================================================================
def should_continue(state: AgentState):
"""Decide next step with robust logic."""
messages = state.get('messages', [])
if not messages:
return "agent"
last_message = messages[-1]
current_turn = state.get('turn', 0)
# Debug: Print what we're checking
msg_type = type(last_message).__name__
print(f"📍 Conditional check - Turn {current_turn}, Last msg type: {msg_type}")
# 1. Check turn limit
if current_turn >= MAX_TURNS:
print(f"🛑 Max turns ({MAX_TURNS}) reached")
return END
# 2. If last message is ToolMessage, agent needs to process it
if isinstance(last_message, ToolMessage):
print(f"📨 Tool result received from '{last_message.name}' → back to agent")
return "agent"
# 3. If last message is AIMessage with tool calls
if isinstance(last_message, AIMessage) and last_message.tool_calls:
# Only check the FIRST tool call, not all of them
first_tool = last_message.tool_calls[0]
tool_name = first_tool.get("name", "")
if tool_name == "final_answer_tool":
return END
else:
return "tools"
# 4. If AIMessage but no tool calls (reasoning text)
if isinstance(last_message, AIMessage) and not last_message.tool_calls:
# Check for consecutive AI messages (loop)
if len(messages) >= 2 and isinstance(messages[-2], AIMessage) and not messages[-2].tool_calls:
print(f"⚠️ Loop detected: 2 consecutive AI messages without tools")
return END
print(f"💭 AI message without tool call → continuing to agent (will force tool)")
return "agent"
# 5. Default: continue to agent
print(f"🔄 Default → continuing to agent")
# =============================================================================
# ENHANCED AGENT CLASS
# =============================================================================
class PlanningReflectionAgent:
def __init__(self):
print("🧠 PlanningReflectionAgent initializing...")
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
if not GROQ_API_KEY:
raise ValueError("GROQ_API_KEY not set!")
HUGGINGFACEHUB_API_TOKEN = os.getenv("HUGGINGFACEHUB_API_TOKEN")
if not HUGGINGFACEHUB_API_TOKEN:
raise ValueError("HUGGINGFACEHUB_API_TOKEN secret is not set! Please add it to your Space secrets.")
GOOGLE_API_KEY = os.getenv("GEMINI_API_KEY")
if not GOOGLE_API_KEY:
raise ValueError("GOOGLE_API_KEY not set!")
self.tools = defined_tools
# Initialize RAG
if not initialize_rag_components():
print("⚠️ RAG components failed to initialize.")
# Build tool descriptions
tool_desc_list = []
for tool in self.tools:
if tool.args_schema:
schema = tool.args_schema.model_json_schema()
args_desc = [f" - {p}: {d.get('description', '')}"
for p, d in schema.get('properties', {}).items()]
desc = f"- {tool.name}:\n {tool.description}\n" + "\n".join(args_desc)
else:
desc = f"- {tool.name}: {tool.description}"
tool_desc_list.append(desc)
tool_descriptions = "\n".join(tool_desc_list)
# ULTRA-AGGRESSIVE SYSTEM PROMPT
self.system_prompt = f"""You are an elite AI agent for GAIA benchmark. Your ONLY job: provide the EXACT answer requested.
═══════════════════════════════════════════════════════════════
⚠️ ABSOLUTE RULES - VIOLATE THESE AND YOU FAIL:
═══════════════════════════════════════════════════════════════
1. **EVERY TURN MUST CALL EXACTLY ONE TOOL** - No exceptions
2. **NEVER OUTPUT REASONING TEXT WITHOUT A TOOL CALL** - You will fail
3. **IDENTIFY QUESTION TYPE FIRST** - Logic? Factual? Data? Math?
4. **LOGIC PUZZLES**: think_through_logic → calculator (if needed) → validate → final_answer
5. **FACTUAL QUESTIONS**: search_tool → validate → final_answer
6. **DATA QUESTIONS**: read_file → code_interpreter → validate → final_answer
7. **ALWAYS VALIDATE**: Call validate_answer() before final_answer_tool()
8. **FINAL ANSWER FORMAT**: EXACTLY what was asked. NO "The answer is..." or explanations
═══════════════════════════════════════════════════════════════
📋 QUESTION TYPE GUIDE:
═══════════════════════════════════════════════════════════════
**RIDDLES/LOGIC PUZZLES** (No web search needed):
- Brain teasers, puzzles, logical deduction
- Strategy: think_through_logic → calculator (if math) → validate → final_answer
- Example: "If 200 coins, 30 face-down, divide into equal piles..."
Turn 1: think_through_logic("Adventurer takes 30 coins and flips them")
Turn 2: calculator("30") [if needed]
Turn 3: validate_answer("30", question)
Turn 4: final_answer_tool("30")
**FACTUAL/RESEARCH** (Need web):
- Who, what, when, where questions
- Strategy: search_tool → scrape_and_retrieve → validate → final_answer
- Example: "What was Einstein's birthplace population in 1900?"
Turn 1: search_tool("Albert Einstein birthplace")
Turn 2: search_tool("Ulm Germany population 1900")
Turn 3: validate_answer("50000", question)
Turn 4: final_answer_tool("50000")
**DATA ANALYSIS** (Need files):
- CSV/Excel questions
- Strategy: list_directory → read_file → code_interpreter → validate → final_answer
**SIMPLE MATH**:
- Calculations
- Strategy: calculator() → validate_answer() → final_answer_tool()
═══════════════════════════════════════════════════════════════
🎓 CRITICAL EXAMPLES:
═══════════════════════════════════════════════════════════════
Example 1: Logic Puzzle
Q: "Coin riddle with 200 coins, 30 face-down..."
✅ CORRECT:
Turn 1: think_through_logic("Take 30 coins, flip all")
Turn 2: validate_answer("30", "coin riddle...")
Turn 3: final_answer_tool("30")
❌ WRONG:
Turn 1: [reasoning text without tool] ← FAILS!
Example 2: Letter Bank Puzzle
Q: "Use letters to spell sentences, which letters need changing?"
✅ CORRECT:
Turn 1: code_interpreter("code to count letters...")
Turn 2: validate_answer("A, B, C", question)
Turn 3: final_answer_tool("A, B, C")
Example 3: Math Problem
Q: "System of equations to solve..."
✅ CORRECT:
Turn 1: code_interpreter("import numpy; solve equations...")
Turn 2: validate_answer("0, 1, 2", question)
Turn 3: final_answer_tool("0, 1, 2")
═══════════════════════════════════════════════════════════════
📚 AVAILABLE TOOLS:
═══════════════════════════════════════════════════════════════
{tool_descriptions}
═══════════════════════════════════════════════════════════════
⚡ EXECUTION RULES:
═══════════════════════════════════════════════════════════════
- If you output text without a tool call, you have FAILED
- If you're unsure, use think_through_logic() to organize thoughts
- ALWAYS call a tool - preferably the right one for the question type
- After EVERY tool result, decide: "Do I have the answer? → validate → submit"
- If stuck after 3 turns: call reflect_on_progress()
REMEMBER: One tool per turn. No reasoning without tools. Exact answer format.
═══════════════════════════════════════════════════════════════
"""
#. Initialize the LLM ()
print("Initializing Groq LLM...")
try:
self.llm_with_tools = ChatGroq(
temperature=0,
groq_api_key=GROQ_API_KEY,
model_name="qwen/qwen3-32b",
max_tokens=4096,
timeout=60
).bind_tools(self.tools, tool_choice="auto")
print("✅ LLM initialized without FORCED tool usage.")
except Exception as e:
print(f"❌ Error initializing HuggingFace: {e}")
raise
print("Initializing LLM Endpoint...")
# print("Initializing HuggingFace LLM...")
#
# llm = HuggingFaceEndpoint(
# repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1", # Free on HF Inference API
# huggingfacehub_api_token=HUGGINGFACEHUB_API_TOKEN,
# max_new_tokens=4096,
# temperature=0.01,
# )
# chat_llm = ChatHuggingFace(llm=llm)
# print("✅ HuggingFace LLM Endpoint initialized.")
#
# # Bind tools to the LLM
# self.llm_with_tools = chat_llm.bind_tools(self.tools)
# print("✅ Tools bound to LLM.")
# print("Initializing Google Gemini LLM...")
# try:
# self.llm_with_tools = ChatGoogleGenerativeAI(
# model="gemini-2.5-flash", # Latest model
# google_api_key=GOOGLE_API_KEY,
# temperature=0,
# max_output_tokens=8192,
# timeout=60,
# convert_system_message_to_human=True # Important for Gemini
# ).bind_tools(self.tools, tool_choice="auto")
# print("✅ Gemini LLM initialized.")
# except Exception as e:
# print(f"❌ Error initializing Gemini: {e}")
# raise
# Agent Node with AGGRESSIVE tool forcing
def agent_node(state: AgentState):
current_turn = state.get('turn', 0) + 1
print(f"\n{'='*70}")
print(f"🤖 AGENT TURN {current_turn}/{MAX_TURNS}")
print('='*70)
if current_turn > MAX_TURNS:
return {
"messages": [SystemMessage(content="Max turns reached.")],
"turn": current_turn
}
# Check if we should force reflection
consecutive_errors = state.get('consecutive_errors', 0)
should_reflect = (current_turn > 5 and current_turn % REFLECT_EVERY_N_TURNS == 0) or consecutive_errors >= 3
messages_to_send = state["messages"].copy()
# Add tool-forcing message if last turn had no tool call
if len(messages_to_send) >= 2:
last_msg = messages_to_send[-1]
if isinstance(last_msg, AIMessage) and not last_msg.tool_calls:
force_msg = SystemMessage(
content="⚠️ CRITICAL: You MUST call a tool this turn. NO reasoning text. Pick the most appropriate tool and call it now."
)
messages_to_send.append(force_msg)
print("🚨 Injecting tool-forcing message")
# Add reflection hint if needed
if should_reflect:
hint = SystemMessage(
content="⚠️ HINT: Multiple turns without progress. Consider calling reflect_on_progress() or try a different approach."
)
messages_to_send.append(hint)
print("🤔 Injecting reflection hint")
# Invoke LLM with retries and fallback
max_retries = 3
ai_message = None
for attempt in range(max_retries):
try:
ai_message = self.llm_with_tools.invoke(messages_to_send)
# If we got a valid response with tool calls, break
if ai_message.tool_calls:
break
# If no tool calls, this is a problem
print(f"⚠️ LLM returned no tool calls on attempt {attempt+1}")
except Exception as e:
error_str = str(e)
print(f"⚠️ LLM attempt {attempt+1}/{max_retries} failed: {error_str[:200]}")
# If tool_use_failed, try without strict binding
if "tool_use_failed" in error_str and attempt < max_retries - 1:
print("🔧 Trying without strict tool enforcement...")
try:
simple_llm = ChatGroq(
temperature=0,
groq_api_key=os.getenv("GROQ_API_KEY"),
model_name="llama-3.3-70b-versatile",
max_tokens=4096,
timeout=60
)
# Add explicit tool forcing to the message
force_tool_msg = SystemMessage(
content="You MUST call a tool. Respond with a tool call, not reasoning text."
)
ai_message = simple_llm.invoke(messages_to_send + [force_tool_msg])
# Try to parse tool calls from content
if ai_message.content and not ai_message.tool_calls:
parsed = parse_tool_call_from_string(ai_message.content, self.tools)
if parsed:
ai_message.tool_calls = parsed
ai_message.content = ""
print("✓ Fallback parsing succeeded")
break
except Exception as e2:
print(f"⚠️ Fallback also failed: {e2}")
if attempt == max_retries - 1:
# Last resort: inject a default tool call
print("🚨 All attempts failed - forcing think_through_logic")
ai_message = AIMessage(
content="",
tool_calls=[ToolCall(
name="think_through_logic",
args={"reasoning": "Processing question"},
id=str(uuid.uuid4())
)]
)
else:
time.sleep(2 ** attempt)
# If still no tool calls after all attempts, force one
if not ai_message.tool_calls:
if isinstance(ai_message.content, str) and ai_message.content.strip():
# Try one more parse
parsed = parse_tool_call_from_string(ai_message.content, self.tools)
if parsed:
ai_message.tool_calls = parsed
ai_message.content = ""
print("✓ Final parse succeeded")
else:
# Absolute last resort
print("🚨 EMERGENCY: Forcing think_through_logic")
ai_message.tool_calls = [ToolCall(
name="think_through_logic",
args={"reasoning": "analyzing question"},
id=str(uuid.uuid4())
)]
ai_message.content = ""
# Track tool usage
tool_history = state.get('tool_history', [])
has_plan = state.get('has_plan', False)
if ai_message.tool_calls:
tool_name = ai_message.tool_calls[0]['name']
print(f"🔧 Tool Call: {tool_name}")
tool_history.append(tool_name)
if tool_name == "create_plan":
has_plan = True
else:
print(f"⚠️ No tool call (this shouldn't happen!)")
print(f"💭 Content: {ai_message.content[:200]}...")
return {
"messages": [ai_message],
"turn": current_turn,
"has_plan": has_plan,
"tool_history": tool_history,
"last_tool_was_thinking": ai_message.tool_calls and ai_message.tool_calls[0]['name'] == 'think_through_logic'
}
# Tool Node with Error Tracking (FIXED)
def tool_node_wrapper(state: AgentState):
"""Executes tools and tracks errors."""
print(f"🔧 Executing tools...")
# Create fresh ToolNode instance
tool_executor = ToolNode(self.tools)
# Invoke properly
result = tool_executor.invoke(state)
# Track errors
consecutive_errors = state.get('consecutive_errors', 0)
if result.get('messages'):
last_msg = result['messages'][-1]
if isinstance(last_msg, ToolMessage):
if "Error" in last_msg.content or "error" in last_msg.content.lower():
consecutive_errors += 1
print(f"⚠️ Tool error detected (consecutive: {consecutive_errors})")
else:
consecutive_errors = 0
result['consecutive_errors'] = consecutive_errors
return result
# Build Graph
print("Building graph...")
graph_builder = StateGraph(AgentState)
graph_builder.add_node("agent", agent_node)
graph_builder.add_node("tools", tool_node_wrapper)
graph_builder.add_edge(START, "agent")
graph_builder.add_conditional_edges(
"agent",
should_continue,
{
"tools": "tools",
"agent": "agent",
END: END
}
)
graph_builder.add_edge("tools", "agent")
self.graph = graph_builder.compile()
print("✅ Graph compiled successfully.")
def __call__(self, question: str, file_path: str = None) -> str:
"""Execute agent on a question."""
print(f"\n{'='*70}")
print(f"🎯 NEW QUESTION")
print(f"{'='*70}")
print(f"Q: {question[:200]}{'...' if len(question) > 200 else ''}")
if file_path:
print(f"📎 File attached: {file_path}")
print(f"{'='*70}\n")
# Enhanced question context with file information
question_text = question
if file_path:
file_ext = Path(file_path).suffix.lower()
file_type = "unknown"
if file_ext in ['.jpg', '.jpeg', '.png', '.gif', '.bmp']:
file_type = "image"
elif file_ext in ['.mp3', '.wav', '.m4a', '.flac']:
file_type = "audio"
elif file_ext in ['.csv', '.xlsx', '.xls']:
file_type = "data"
elif file_ext in ['.txt', '.pdf', '.doc', '.docx']:
file_type = "document"
question_text += f"\n\n[FILE ATTACHED: {file_path}]"
question_text += f"\n[FILE TYPE: {file_type}]"
question_text += f"\nIMPORTANT: Use the appropriate tool to access this file first!"
graph_input = {
"messages": [
SystemMessage(content=self.system_prompt),
HumanMessage(content=question_text)
],
"file_path": file_path,
"turn": 0,
"has_plan": False,
"consecutive_errors": 0,
"tool_history": [],
"last_tool_was_thinking": False
}
final_answer = "AGENT FAILED TO PRODUCE ANSWER"
all_messages = []
try:
config = {"recursion_limit": MAX_TURNS + 10}
for event in self.graph.stream(graph_input, stream_mode="values", config=config):
if not event.get('messages'):
continue
all_messages = event["messages"]
last_message = all_messages[-1]
# Check for final answer
if isinstance(last_message, AIMessage) and last_message.tool_calls:
for tool_call in last_message.tool_calls:
if tool_call.get("name") == "final_answer_tool":
args = tool_call.get('args', {})
if 'answer' in args:
final_answer = args['answer']
print(f"\n{'='*70}")
print(f"✅ FINAL ANSWER: '{final_answer}'")
print(f"{'='*70}\n")
break
elif isinstance(last_message, ToolMessage):
preview = last_message.content[:200].replace('\n', ' ')
print(f"📊 Tool '{last_message.name}' result: {preview}...")
elif isinstance(last_message, AIMessage) and not last_message.tool_calls:
print(f"💭 AI: {last_message.content[:200]}...")
# If no final answer, try to extract from tool messages
if final_answer == "AGENT FAILED TO PRODUCE ANSWER":
print("⚠️ No final_answer_tool called. Checking tool results...")
for msg in reversed(all_messages):
if isinstance(msg, ToolMessage):
if msg.name in ["calculator", "think_through_logic", "code_interpreter"]:
content = msg.content.strip()
# Look for short, answer-like content
if content and len(content) < 200 and not content.startswith("Error"):
# Extract just the result part
lines = content.split('\n')
for line in reversed(lines):
if line.strip() and not line.startswith(('✅', '⚠️', 'Next', 'Remember')):
final_answer = line.strip()
print(f"📝 Extracted from {msg.name}: '{final_answer}'")
break
break
# Clean the answer
cleaned = str(final_answer).strip()
# Remove prefixes
prefixes = [
"the answer is:", "here is the answer:", "based on",
"final answer:", "answer:", "the final answer is:",
"my answer is:", "according to", "i found that",
"the result is:", "result:"
]
for prefix in prefixes:
if cleaned.lower().startswith(prefix.lower()):
potential = cleaned[len(prefix):].strip()
if potential:
cleaned = potential
break
# Remove code fences and quotes
cleaned = remove_fences_simple(cleaned)
while cleaned.startswith("`") and cleaned.endswith("`"):
cleaned = cleaned[1:-1].strip()
if (cleaned.startswith('"') and cleaned.endswith('"')) or \
(cleaned.startswith("'") and cleaned.endswith("'")):
cleaned = cleaned[1:-1].strip()
# Remove trailing period for short answers
if cleaned.endswith('.') and len(cleaned.split()) < 10:
cleaned = cleaned[:-1]
print(f"\n{'='*70}")
print(f"🎉 RETURNING ANSWER")
print(f"{'='*70}")
print(f"{cleaned}")
print(f"{'='*70}\n")
return cleaned
except Exception as e:
print(f"❌ Graph error: {e}")
print(traceback.format_exc())
return f"AGENT ERROR: {e}"
# =============================================================================
# GLOBAL AGENT INSTANTIATION
# =============================================================================
agent = None
try:
initialize_rag_components()
agent = PlanningReflectionAgent()
print("✅ Global PlanningReflectionAgent instantiated.")
# Verify it's callable
if not callable(agent):
print("❌ ERROR: Agent not callable!")
agent = None
else:
print("✅ Agent is callable.")
if asr_pipeline is None:
print("⚠️ ASR Pipeline not loaded.")
except Exception as e:
print(f"❌ FATAL: Agent initialization failed: {e}")
traceback.print_exc()
agent = None
# =============================================================================
# RUN AND SUBMIT FUNCTION
# =============================================================================
def run_and_submit_all(profile: gr.OAuthProfile | None):
"""
Fetches all questions, runs the BasicAgent on them, submits all answers,
and displays the results.
"""
space_id = os.getenv("SPACE_ID")
if profile:
username = f"{profile.username}"
print(f"User logged in: {username}")
else:
print("User not logged in.")
return "Please Login to Hugging Face with the button.", None
# Use the globally instantiated agent
global agent
if agent is None:
error_msg = "FATAL: Agent failed to initialize at startup. Check logs for errors."
print(error_msg)
return error_msg, None
print("✅ Using globally instantiated PlanningReflectionAgent")
api_url = DEFAULT_API_URL
questions_url = f"{api_url}/questions"
submit_url = f"{api_url}/submit"
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
print(agent_code)
# 2. Fetch Questions
print(f"\n{'='*70}")
print(f"📥 FETCHING QUESTIONS")
print(f"{'='*70}")
print(f"Fetching questions from: {questions_url}")
try:
response = requests.get(questions_url, timeout=15)
response.raise_for_status()
questions_data = response.json()
if not questions_data:
print("Fetched questions list is empty.")
return "Fetched questions list is empty or invalid format.", None
print(f"✅ Fetched {len(questions_data)} questions.")
print(f"{'='*70}\n")
except requests.exceptions.RequestException as e:
print(f"❌ Error fetching questions: {e}")
return f"Error fetching questions: {e}", None
except requests.exceptions.JSONDecodeError as e:
print(f"❌ Error decoding JSON response from questions endpoint: {e}")
print(f"Response text: {response.text[:500]}")
return f"Error decoding server response for questions: {e}", None
except Exception as e:
print(f"❌ An unexpected error occurred fetching questions: {e}")
return f"An unexpected error occurred fetching questions: {e}", None
# Load answer sheet
answer_sheet = load_answer_sheet("answer_sheet_json.json")
# If answer sheet doesn't exist, create template
if not answer_sheet:
create_answer_sheet_template(questions_data, "answer_sheet.json")
print("\n⚠️ Please fill in the answer_sheet.json file with correct answers")
print(" Then run the script again to check agent performance\n")
results = []
local_correct = 0
local_total = 0
# 3. Run your Agent
print(f"\n{'='*70}")
print(f"🚀 STARTING EVALUATION")
print(f"{'='*70}")
print(f"Total questions to process: {len(questions_data)}")
print(f"{'='*70}\n")
results_log = []
answers_payload = []
for idx, item in enumerate(questions_data, 1):
print(f"\n{'='*70}")
print(f"📝 PROCESSING QUESTION {idx}/{len(questions_data)}")
print(f"{'='*70}")
task_id = item.get("task_id")
question_text = item.get("question")
correct_answer = answer_sheet.get(task_id, "")
# Look for file locally in files/ directory
local_file_path = None
files_dir = "files"
try:
# Check if files directory exists
if os.path.exists(files_dir):
# Look for any file that starts with the task_id
matching_files = [f for f in os.listdir(files_dir) if f.startswith(task_id)]
if matching_files:
# Use the first matching file
local_file_path = os.path.join(files_dir, matching_files[0])
file_size = os.path.getsize(local_file_path)
abs_path = os.path.abspath(local_file_path)
print(f"✅ Found file: {matching_files[0]} ({file_size} bytes)")
print(f" Path: {abs_path}")
else:
print(f"ℹ️ No file found for task {task_id}, proceeding without file.")
else:
print(f"⚠️ Warning: '{files_dir}' directory not found.")
except Exception as e:
print(f"❌ Error looking for file: {e}")
try:
# Pass file_path to agent
submitted_answer = agent(question_text, local_file_path)
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
# Check if answer is correct
is_correct = submitted_answer.strip().lower() == correct_answer.strip().lower()
correctness = "✅ CORRECT" if is_correct else "❌ WRONG"
# Log with correctness indicator
print(f"\n{correctness} - Task {task_id}")
print(f" Submitted: '{submitted_answer}'")
print(f" Expected: '{correct_answer}'")
results_log.append({
"Task ID": task_id,
"Question": question_text[:100] + "..." if len(question_text) > 100 else question_text,
"Submitted Answer": submitted_answer,
"Correct Answer": correct_answer,
"Status": "✅" if is_correct else "❌"
})
print(f"✅ Question {idx}/{len(questions_data)} completed")
except Exception as e:
print(f"❌ Error running agent on task {task_id}: {e}")
print(traceback.format_exc())
results_log.append({
"Task ID": task_id,
"Question": question_text[:100] + "..." if len(question_text) > 100 else question_text,
"Submitted Answer": f"AGENT ERROR: {e}",
"Correct Answer": correct_answer,
"Status": "❌"
})
# Continue with other questions even if one fails
answers_payload.append({"task_id": task_id, "submitted_answer": f"ERROR: {str(e)[:100]}"})
# Summary after all questions processed
print(f"\n{'='*70}")
print(f"✅ ALL QUESTIONS PROCESSED")
print(f"{'='*70}")
print(f"Total answers collected: {len(answers_payload)}")
# Calculate pre-submission accuracy
correct_count = sum(1 for log in results_log if log.get("Status") == "✅")
total_count = len(results_log)
accuracy = (correct_count / total_count * 100) if total_count > 0 else 0
print(f"\n{'='*70}")
print(f"📊 PRE-SUBMISSION SUMMARY")
print(f"{'='*70}")
print(f"Correct: {correct_count}/{total_count} ({accuracy:.1f}%)")
print(f"{'='*70}\n")
if not answers_payload:
print("⚠️ Agent did not produce any answers to submit.")
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
# 4. Prepare Submission
submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
# 5. Submit
print(f"\n{'='*70}")
print(f"📤 SUBMITTING TO API")
print(f"{'='*70}")
print(f"URL: {submit_url}")
print(f"Username: {username}")
print(f"Answers to submit: {len(answers_payload)}")
print(f"{'='*70}\n")
try:
print("⏳ Sending POST request...")
response = requests.post(submit_url, json=submission_data, timeout=60)
print(f"✅ Got response: Status {response.status_code}")
response.raise_for_status()
result_data = response.json()
print(f"\n{'='*70}")
print(f"📊 SUBMISSION RESULTS")
print(f"{'='*70}")
print(f"Response data: {result_data}")
print(f"{'='*70}\n")
final_status = (
f"Submission Successful!\n"
f"User: {result_data.get('username')}\n"
f"Overall Score: {result_data.get('score', 'N/A')}% "
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
f"Message: {result_data.get('message', 'No message received.')}"
)
print(final_status)
print("="*70)
print("✅ Submission successful.")
results_df = pd.DataFrame(results_log)
return final_status, results_df
except requests.exceptions.HTTPError as e:
error_detail = f"Server responded with status {e.response.status_code}."
try:
error_json = e.response.json()
error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
except requests.exceptions.JSONDecodeError:
error_detail += f" Response: {e.response.text[:500]}"
status_message = f"Submission Failed: {error_detail}"
print(f"\n{'='*70}")
print(f"❌ SUBMISSION FAILED")
print(f"{'='*70}")
print(status_message)
print(f"{'='*70}\n")
results_df = pd.DataFrame(results_log)
return status_message, results_df
except requests.exceptions.Timeout:
status_message = "Submission Failed: The request timed out."
print(f"\n{'='*70}")
print(f"❌ SUBMISSION FAILED")
print(f"{'='*70}")
print(status_message)
print(f"{'='*70}\n")
results_df = pd.DataFrame(results_log)
return status_message, results_df
except requests.exceptions.RequestException as e:
status_message = f"Submission Failed: Network error - {e}"
print(f"\n{'='*70}")
print(f"❌ SUBMISSION FAILED")
print(f"{'='*70}")
print(status_message)
print(f"{'='*70}\n")
results_df = pd.DataFrame(results_log)
return status_message, results_df
except Exception as e:
status_message = f"An unexpected error occurred during submission: {e}"
print(f"\n{'='*70}")
print(f"❌ SUBMISSION FAILED")
print(f"{'='*70}")
print(status_message)
print(traceback.format_exc())
print(f"{'='*70}\n")
results_df = pd.DataFrame(results_log)
return status_message, results_df
# --- Build Gradio Interface using Blocks ---
with gr.Blocks() as demo:
gr.Markdown("# Basic Agent Evaluation Runner")
gr.Markdown(
"""
**Instructions:**
1. Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
2. Log in to your Hugging Face account using the button below. This uses your HF username for submission.
3. Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.
---
**Disclaimers:**
Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
"""
)
gr.LoginButton()
run_button = gr.Button("Run Evaluation & Submit All Answers")
status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
run_button.click(
fn=run_and_submit_all,
outputs=[status_output, results_table]
)
if __name__ == "__main__":
print("\n" + "-"*30 + " App Starting " + "-"*30)
space_host_startup = os.getenv("SPACE_HOST")
space_id_startup = os.getenv("SPACE_ID")
if space_host_startup:
print(f"✅ SPACE_HOST found: {space_host_startup}")
print(f" Runtime URL should be: https://{space_host_startup}.hf.space")
else:
print("ℹ️ SPACE_HOST environment variable not found (running locally?).")
if space_id_startup:
print(f"✅ SPACE_ID found: {space_id_startup}")
print(f" Repo URL: https://huggingface.co/spaces/{space_id_startup}")
print(f" Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
else:
print("ℹ️ SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")
print("-"*(60 + len(" App Starting ")) + "\n")
print("Launching Gradio Interface for Basic Agent Evaluation...")
demo.launch(debug=True, share=False)