AK47-M4A4's picture
v1
ae1d0b9
# pylint: disable=no-member
import base64
import gc
import math
import mimetypes
import multiprocessing
import os
import re
import tempfile
import time
import uuid
from datetime import timedelta
from typing import Dict, List, Optional, TypedDict, Union
from urllib.parse import urlparse
import cv2
import imageio
import pandas as pd
import pytesseract
import requests
import torch
import whisper
import yt_dlp
from bs4 import BeautifulSoup, Tag
from dotenv import load_dotenv
from duckduckgo_search import DDGS
from langchain_core.messages import HumanMessage
from langchain_core.tools import tool
from langchain_ollama import ChatOllama
from PIL import Image
from playwright.sync_api import sync_playwright
from youtube_transcript_api import (
NoTranscriptFound,
TranscriptsDisabled,
YouTubeTranscriptApi,
)
load_dotenv()
base_url = os.getenv("OLLAMA_BASE_URL")
model_vision = ChatOllama(
model="gemma3:latest",
base_url=base_url,
)
model_text = ChatOllama(
model="hf.co/lmstudio-community/Qwen2.5-14B-Instruct-GGUF:Q6_K", base_url=base_url
)
@tool
def use_vision_model(question: str) -> str:
"""
A multimodal reasoning model that combines image and text input to answer
questions using the image.
"""
# Extract image paths
image_paths = re.findall(r"[\w\-/\.]+\.(?:png|jpg|jpeg|webp)", question)
image_paths = [p for p in image_paths if os.path.exists(p)]
if not image_paths:
return "No valid image file found in the question."
image_path = image_paths[0]
# # Preprocess the image using OpenCV
# image = cv2.imread(image_path)
# gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# gray = cv2.convertScaleAbs(gray, alpha=1.2, beta=20)
# gray = cv2.GaussianBlur(gray, (5, 5), 0)
# edges = cv2.Canny(gray, 50, 150, apertureSize=3)
# # Create a temporary file for the processed image
# with tempfile.NamedTemporaryFile(suffix=".png", delete=True) as tmp_file:
# temp_image_path = tmp_file.name
# cv2.imwrite(temp_image_path, image)
# Encode the temp image(this code was under with tempfile)
mime_type, _ = mimetypes.guess_type(image_path)
mime_type = mime_type or "image/png"
with open(image_path, "rb") as f:
encoded = base64.b64encode(f.read()).decode("utf-8")
# Prepare the prompt and image for the model
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": question},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{encoded}"},
},
],
}
]
# Invoke the vision model
response = model_vision.invoke(messages)
# Clean up
del messages, encoded, image_path
gc.collect()
torch.cuda.empty_cache()
return str(response.content) if hasattr(response, "content") else str(response)
# YouTube Video Review Tool
@tool
def review_youtube_video(url: str) -> str:
"""Reviews a YouTube video and answers a specific question about that video.
Args:
url (str): the URL to the YouTube video.
question (str): The question you are asking about the video.
Returns:
str: The answer to the question
"""
# Extract video ID from URL (assuming it is in the format https://youtube.com/watch?v=VIDEO_ID)
video_id = url.split("v=")[1]
transcript_url = (
f"https://www.youtube.com/api/timedtext?v={video_id}" # Getting transcript data
)
response = requests.get(transcript_url, timeout=200)
transcript = response.text # This is the transcript (XML or SRT format)
# Prepare the content (just the transcript, no question needed)
transcript_content = f"Here is the transcript of the video: {transcript}"
# Return the transcript content so the main LLM can handle question generation
return transcript_content
# YouTube Frames to Images Tool
@tool
def video_frames_to_images(
url: str,
sample_interval_seconds: int = 5,
) -> List[str]:
"""Extracts frames from a video at specified intervals and saves them as images.
Args:
url (str): the URL to the video.
folder_name (str): the name of the folder to save the images to.
sample_interval_seconds (int): the interval between frames to sample.
Returns:
List[str]: A list of paths to the saved image files.
"""
folder_name = "./frames"
# Create a subdirectory for the frames
frames_dir = os.path.join(folder_name, "frames")
os.makedirs(frames_dir, exist_ok=True)
ydl_opts = {
"format": "bestvideo[height<=1080]+bestaudio/best[height<=1080]/best",
"outtmpl": os.path.join(folder_name, "video.%(ext)s"),
"quiet": True,
"noplaylist": True,
"merge_output_format": "mp4",
"force_ipv4": True,
}
info_extracted = []
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
info = ydl.extract_info(url, download=True)
info_extracted.append(info)
video_path = next(
(
os.path.join(folder_name, f)
for f in os.listdir(folder_name)
if f.endswith(".mp4")
),
None,
)
if not video_path:
raise RuntimeError("Failed to download video as mp4")
reader = imageio.get_reader(video_path)
# metadata = reader.get_meta_data()
fps = 25
duration_seconds = 120
frame_interval = int(fps * sample_interval_seconds)
num_frames = int(fps * duration_seconds)
# if num_frames is None or math.isinf(num_frames):
# num_frames = int(fps * duration_seconds)
# Handle case where the number of frames is infinite or invalid
# if num_frames == float("inf") or not isinstance(num_frames, int):
# reader.close()
# raise RuntimeError("Invalid video length (infinite or not an integer)")
image_paths: List[str] = []
for idx in range(num_frames):
if idx % frame_interval == 0:
# Save frame as image
frame = reader.get_data(idx)
image_path = os.path.join(frames_dir, f"frame_{idx:06d}.jpg")
imageio.imwrite(image_path, frame)
image_paths.append(image_path)
reader.close()
return image_paths
# File Reading Tool
@tool
def read_file(filepath: str) -> str:
"""Reads the content of a PYTHON file.
Args:
filepath (str): the path to the file to read.
Returns:
str: The content of the file.
"""
try:
with open(filepath, "r", encoding="utf-8") as file:
content = file.read()
# Calculate metadata for the prompt
filename = os.path.basename(filepath)
line_count = content.count("\\n") + 1
code_str = content.strip()
# Compose the prompt
prompt = f"""
You are a Python expert and code reviewer. Analyze the following Python script and answer the question provided.
Give Final Answer: the output of the code
Script Length: {line_count} lines
Filename: {filename}
Python Code:
```python
{code_str}
```
"""
model = model_text
# Call the model
message = HumanMessage(content=prompt)
response = model.invoke([message])
torch.cuda.empty_cache()
gc.collect()
# Return the result
if hasattr(response, "content") and isinstance(response.content, str):
return response.content
return str(response)
except FileNotFoundError:
return f"File not found: {filepath}"
except IOError as e:
return f"Error reading file: {str(e)}"
# To run python code
def execute_code(code: str):
"""Helper function to execute the code in a separate process."""
try:
exec(code)
except Exception as e:
raise RuntimeError(f"Error executing the code: {str(e)}") from e
@tool
def run_code_from_file(file_path: str, timeout: int = 10):
"""
Reads a Python file and executes it, with timeout handling.
Args:
file_path (str): The full path to the Python file to execute.
timeout (int): The timeout in seconds before forcefully stopping the execution.
"""
# Check if the file exists
if not os.path.exists(file_path):
raise FileNotFoundError(f"The file {file_path} does not exist.")
# Read the file and get the code to execute
with open(file_path, "r", encoding="utf-8") as file:
code = file.read()
# Start a process to execute the code
process = multiprocessing.Process(target=execute_code, args=(code,))
process.start()
# Wait for the process to finish or timeout
process.join(timeout)
# If the process is still alive after the timeout, terminate it
if process.is_alive():
process.terminate() # Stop the execution
raise TimeoutError(
f"The code execution took longer than {timeout} seconds and was terminated."
)
# File Download Tool
@tool
def download_file_from_url(url: str, directory: str) -> Dict[str, Union[str, None]]:
"""Downloads a file from a URL and saves it to a directory.
Args:
url (str): the URL to download the file from.
directory (str): the directory to save the file to.
Returns:
Dict[str, Union[str, None]]: A dictionary containing the file type and path.
"""
response = requests.get(url, stream=True, timeout=10)
response.raise_for_status()
content_type = response.headers.get("content-type", "").lower()
# Try to get filename from headers
filename = None
cd = response.headers.get("content-disposition", "")
match = re.search(r"filename\*=UTF-8\'\'(.+)", cd) or re.search(
r'filename="?([^"]+)"?', cd
)
if match:
filename = match.group(1)
# If not in headers, try URL
if not filename:
filename = os.path.basename(url.split("?")[0])
# Fallback to generated filename
if not filename:
extension = {
"image/jpeg": ".jpg",
"image/png": ".png",
"image/gif": ".gif",
"audio/wav": ".wav",
"audio/mpeg": ".mp3",
"video/mp4": ".mp4",
"text/plain": ".txt",
"text/csv": ".csv",
"application/vnd.openxmlformats-officedocument.spreadsheetml.sheet": ".xlsx",
"application/vnd.ms-excel": ".xls",
"application/octet-stream": ".bin",
}.get(content_type, ".bin")
filename = f"downloaded_file{extension}"
os.makedirs(directory, exist_ok=True)
file_path = os.path.join(directory, filename)
print(file_path)
with open(file_path, "wb") as f:
for chunk in response.iter_content(chunk_size=8192):
f.write(chunk)
# shutil.copy(file_path, os.getcwd())
return {
"type": content_type,
"filename": filename,
"path": file_path,
}
# Text Extraction from Image Tool
@tool
def extract_text_from_image(image_path: str) -> str:
"""Extracts text from an image using OCR.
Args:
image_path (str): the path to the image to extract text from.
Returns:
str: The text extracted from the image.
"""
image = Image.open(image_path)
text = pytesseract.image_to_string(image)
return f"Extracted text from image:\n\n{text}"
# CSV Analysis Tool
@tool
def analyze_csv_file(file_path: str, query: str) -> str:
"""Analyzes a CSV file and answers questions about its contents using an
Ollama model.
Args:
file_path (str): The path to the CSV file to analyze.
query (str): The question to answer about the CSV file.
Returns:
str: The result of the analysis.
"""
# Load the CSV file
df = pd.read_csv(file_path)
df_str = df.to_string(index=False)
# Compose the prompt
prompt = f"""
You are a data analyst. Analyze the following CSV data and answer the question provided.
CSV Dimensions: {df.shape[0]} rows × {df.shape[1]} columns
CSV Data:
{df_str}
Please provide:
1. A summary of the data structure and content
2. Key patterns and insights
3. Potential data quality issues
4. Suggestions for analysis
User Query:
{query}
Format your response in markdown with sections and bullet points.
"""
model = model_text
# Call the model
response = model.invoke([{"type": "text", "text": prompt}])
del df
torch.cuda.empty_cache()
gc.collect()
# Return the result
if hasattr(response, "content") and isinstance(response.content, str):
return response.content
return str(response)
# Excel Analysis Tool
@tool
def analyze_excel_file(file_path: str) -> str:
"""Analyzes an Excel file and answers questions about its contents using an
Ollama model
Args:
file_path (str): the path to the Excel file to analyze.
query (str): the question to answer about the Excel file.
Returns:
str: The result of the analysis.
"""
llm = model_text
print(file_path)
# Read all sheets from the Excel file
excel_file = pd.ExcelFile(file_path)
sheet_names = excel_file.sheet_names
result = f"Excel file loaded with {len(sheet_names)} sheets: {', '.join(sheet_names)}\n\n"
for sheet_name in sheet_names:
df = pd.read_excel(file_path, sheet_name=sheet_name)
df_str = df.to_string()
# Build the prompt
prompt = f"""Analyze the following Excel sheet data and answer the user's query.
Sheet Name: {sheet_name}
Dimensions: {len(df)} rows × {len(df.columns)} columns
Data:
{df_str}
Please provide:
1. A summary of the data structure and content
2. List all the values of the columns in a proper table format.
3. If a file contains food items, assume it refers to the
monetary value of the items, not the quantity sold.
4. If the File contains food items, make a new list which
contains the name of all the food item in the column only (not including drinks).
5. If the file contains any time of monetary value its in USD with two decimal places.
Format the response clearly using headings and bullet points."""
# Call the LLM with the prompt
response = llm.invoke([HumanMessage(content=prompt)])
result += f"=== Sheet: {sheet_name} ===\n"
result += str(response.content) + "\n"
result += "=" * 50 + "\n\n"
del df
gc.collect()
excel_file.close()
torch.cuda.empty_cache()
return result
# Audio Transcription Tool
def transcribe_audio(audio_file_path: str) -> str:
"""Transcribes an audio file using Whisper's audio capabilities.
Always give Final Answer of the question in a specific format for example list all the pages mentioned in increasing order in one line.
Change vanilla extract to pure vanilla extract in the final answer.
Args:
audio_file_path (str): The path to the audio file to transcribe.
mime_type (str): The MIME type of the audio file.
Returns:
str: The transcript of the audio file.
Raises:
ValueError: If the MIME type is not supported.
"""
model = whisper.load_model("base")
result = model.transcribe(audio_file_path)
assert isinstance(result["text"], str)
del model
torch.cuda.empty_cache()
gc.collect()
return result["text"]
def _extract_video_id(url: str) -> Optional[str]:
"""Extract video ID from YouTube URL.
Args:
url (str): the URL to the YouTube video.
Returns:
str: The video ID of the YouTube video.
"""
patterns = [
r"(?:youtube\.com\/watch\?v=|youtube\.com\/embed\/|youtu\.be\/)([^&\n?#]+)",
r"(?:youtube\.com\/v\/|youtube\.com\/e\/|youtube\.com\/user\/[^\/]+\/|youtube\.com\/[^\/]+\/|youtube\.com\/embed\/|youtu\.be\/)([^&\n?#]+)",
]
for pattern in patterns:
match = re.search(pattern, url)
if match:
return match.group(1)
return None
@tool
def transcribe_youtube(url: str) -> str:
"""
Transcribes a YouTube video using YouTube Transcript API or ChatOllama with Whisper as fallback.
This function first tries to fetch the transcript of a YouTube video using the YouTube Transcript API.
If the transcript is unavailable (e.g., due to captions being disabled), it falls back to using
ChatOllama integrated with Whisper to transcribe the audio.
Args:
url (str): The URL to the YouTube video.
Returns:
str: The transcript of the YouTube video, or an error message if transcription fails.
"""
try:
# Try using YouTube Transcript API
video_id = _extract_video_id(url)
transcript = ""
transcript_chunks = YouTubeTranscriptApi.get_transcript(
video_id, languages=["en"]
)
for chunk in transcript_chunks:
timestamp = str(timedelta(seconds=int(chunk["start"])))
transcript += f"[{timestamp}] {chunk['text']}\n"
# Return API transcript if available
if transcript.strip():
return transcript
except (TranscriptsDisabled, NoTranscriptFound, Exception) as err:
try:
with tempfile.TemporaryDirectory() as tmpdir:
# Download audio from YouTube
ydl_opts = {
"format": "bestaudio/best",
"outtmpl": os.path.join(tmpdir, "audio.%(ext)s"),
"quiet": True,
"noplaylist": True,
"postprocessors": [
{
"key": "FFmpegExtractAudio",
"preferredcodec": "wav",
"preferredquality": "192",
}
],
}
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
info = ydl.extract_info(url, download=True)
if info is not None:
title = info.get("title", "Unknown Title") # Type:None
duration = info.get("duration", 0) # in seconds
uploader = info.get("uploader", "Unknown Uploader")
else:
title = "Unknown Title"
duration = 0
uploader = "Unknown Uploader"
audio_path = next(
(
os.path.join(tmpdir, f)
for f in os.listdir(tmpdir)
if f.endswith(".wav")
),
None,
)
if not audio_path:
raise RuntimeError("Failed to download or convert audio") from err
# Use Whisper for initial transcription
whisper_model = whisper.load_model("base")
transcription = whisper_model.transcribe(audio_path, verbose=False)
raw_transcript = transcription["text"]
del whisper_model
gc.collect()
torch.cuda.empty_cache()
result = f"Title: {title}\nUploader: {uploader}\nDuration: {duration} seconds\nTranscript: {raw_transcript}"
return result
except Exception as fallback_exc:
raise RuntimeError("Fallback Transcription failed") from fallback_exc
return "Transcription failed unexpectedly."
@tool
def website_scrape(url: str) -> str:
"""scrapes a website and returns the text.
args:
url (str): the url to the website to scrape.
returns:
str: the text of the website.
"""
try:
parsed_url = urlparse(url)
if not parsed_url.scheme or not parsed_url.netloc:
raise ValueError(
f"Invalid URL: '{url}'. Call `duckduckgo_search` first to get a valid URL."
)
with sync_playwright() as p:
browser = p.chromium.launch(headless=True)
page = browser.new_page()
page.goto(url, wait_until="networkidle", timeout=60000)
page.wait_for_load_state("domcontentloaded")
html_content = page.content()
browser.close()
soup = BeautifulSoup(html_content, "html.parser")
relevant_text = ""
# for header in soup.find_all(["h2", "h3"]):
# heading_text = header.get_text().strip().lower()
# if "discography" in heading_text or "studio albums" in heading_text:
# section_texts = []
# tag = header.find_next_sibling()
# while tag and (
# not isinstance(tag, Tag) or tag.name not in ["h2", "h3"]
# ):
# section_texts.append(tag.get_text(separator=" ", strip=True))
# tag = tag.find_next_sibling()
# relevant_text = "\n\n".join(section_texts)
# break
# if not relevant_text:
# article = soup.find("article")
# if article:
# relevant_text = article.get_text(separator=" ", strip=True)
# if not relevant_text:
relevant_text = soup.get_text(separator=" ", strip=True)
# step 2: chunk the text (optional but recommended)
def chunk_text(text, max_length=1000):
words = text.split()
chunks = []
for i in range(0, len(words), max_length):
chunks.append(" ".join(words[i : i + max_length]))
return chunks
chunks = chunk_text(relevant_text)
# return only the first 2–3 chunks to keep it concise
return "\n\n".join(chunks[:5])
except ValueError as e:
# Catch URL validation errors
return str(e)
except Exception as e:
# Catch other unexpected errors
return f"Scraping failed: {str(e)}"
class SearchResult(TypedDict):
query: str
status: str
attempt: int
results: Optional[List[dict]]
error: Optional[str]
@tool
def duckduckgo_search(query: str, max_results: int = 10) -> SearchResult:
"""
Perform a DuckDuckGo search with retry and backoff.
Use this FIRST before invoking and scraping tools.
Args:
query: The search query string.
max_results: Max number of results to return (default 10).
Returns:
A dict with the query, results, status, attempt count, and any error.
"""
max_retries = 3
base_delay = 2
backoff_factor = 2
for attempt in range(max_retries):
try:
with DDGS() as ddgs:
results = ddgs.text(keywords=query, max_results=max_results)
if results:
formatted_results = [
{
"title": result.get("title", ""),
"url": result.get("href", ""),
"body": result.get("body", ""),
}
for result in results
]
return {
"query": query,
"status": "success",
"attempt": attempt + 1,
"results": formatted_results,
"error": None,
}
except Exception as e:
print(f"[DuckDuckGo Tool] Attempt {attempt + 1} failed: {e}")
time.sleep(base_delay * (backoff_factor**attempt))
return {
"query": query,
"status": "failed",
"attempt": max_retries,
"results": None,
"error": "Max retries exceeded or request failed.",
}
@tool
def reverse_decoder(question: str) -> str:
"""Decodes a reversed sentence if the input appears to be written backward.
Args:
question (str): The possibly reversed question string.
Returns:
str: The decoded sentence.
"""
# Remove leading punctuation if present
cleaned = question.strip().strip(".!?")
# Check if it's likely reversed (simple heuristic: mostly lowercase, reversed word order)
reversed_text = cleaned[::-1]
return reversed_text