INTRODUCTION TO AGENTIC AI FOR DOCUMENT CREATION
An Agentic AI system represents a sophisticated approach to artificial intelligence where autonomous agents can perform complex, multi-step tasks with minimal human intervention. In the context of PowerPoint creation, such a system transforms a simple user prompt about a topic into a comprehensive presentation by orchestrating multiple AI components working in concert.
The core concept revolves around creating an intelligent agent that can understand user requirements, conduct research, process information, and generate professional presentations automatically. This system goes beyond simple template filling by incorporating real-time research capabilities, advanced content processing, and intelligent design decisions.
The agent operates through a series of coordinated steps that mirror how a human researcher and presenter would approach the task. It begins by understanding the user's topic, searches for relevant information across the internet, downloads and processes documents, extracts meaningful content, organizes information logically, and finally creates visually appealing slides that follow established UX principles.
SYSTEM ARCHITECTURE OVERVIEW
The architecture of our Agentic AI system follows a modular design where each component has specific responsibilities while maintaining clear interfaces for communication. The system consists of several interconnected modules that work together to transform user input into polished presentations.
At the highest level, we have the Orchestrator Agent that coordinates all activities. This agent receives user prompts and manages the workflow through various specialized components. The Web Search Agent handles internet research and document discovery. The Content Processor extracts and cleans text from downloaded documents. The RAG System provides intelligent information retrieval and synthesis. The GraphRAG component creates knowledge graphs and ontologies. The Visual Processor extracts images and charts using Vision Language Models. Finally, the Presentation Generator creates the actual PowerPoint files.
Each component operates independently but communicates through well-defined APIs. This modular approach allows for easy maintenance, testing, and future enhancements. The system also includes a Configuration Manager that handles user preferences such as PowerPoint themes and presentation styles.
Here's a foundational code example that demonstrates the basic orchestrator structure:
The following code example shows how to implement the main orchestrator class that coordinates all system components. This class serves as the central hub that manages the workflow from user input to final presentation generation.
import asyncio
import logging
from typing import Dict, List, Optional
from dataclasses import dataclass
from pathlib import Path
@dataclass
class UserRequest:
topic: str
theme: Optional[str] = None
max_slides: int = 10
output_folder: str = "./presentations"
class AgenticPowerPointSystem:
def __init__(self, config_path: str):
self.config = self.load_configuration(config_path)
self.web_searcher = WebSearchAgent(self.config)
self.content_processor = ContentProcessor(self.config)
self.rag_system = RAGSystem(self.config)
self.graph_rag = GraphRAGSystem(self.config)
self.visual_processor = VisualProcessor(self.config)
self.presentation_generator = PresentationGenerator(self.config)
self.logger = logging.getLogger(__name__)
async def create_presentation(self, request: UserRequest) -> str:
"""Main orchestration method for presentation creation"""
self.logger.info(f"Starting presentation creation for topic: {request.topic}")
# Step 1: Search and download documents
documents = await self.web_searcher.search_and_download(
request.topic,
max_documents=20
)
# Step 2: Extract and process content
processed_content = await self.content_processor.process_documents(documents)
# Step 3: Build RAG system with processed content
await self.rag_system.index_content(processed_content)
# Step 4: Create knowledge graph and ontology
ontology = await self.graph_rag.create_ontology(processed_content)
# Step 5: Extract visual elements
visual_elements = await self.visual_processor.extract_visuals(documents)
# Step 6: Generate presentation structure
presentation_structure = await self.generate_presentation_structure(
request, processed_content, ontology
)
# Step 7: Create PowerPoint file
output_path = await self.presentation_generator.create_presentation(
presentation_structure,
visual_elements,
request.theme,
request.output_folder
)
self.logger.info(f"Presentation created successfully: {output_path}")
return output_path
This orchestrator demonstrates how the system coordinates multiple agents to accomplish the complex task of presentation creation. Each method call represents a significant subsystem that we'll explore in detail throughout this article.
WEB SEARCH AND DOCUMENT DISCOVERY COMPONENT
The Web Search Agent serves as the research arm of our system, responsible for finding relevant documents across the internet based on user-specified topics. This component must be intelligent enough to formulate effective search queries, evaluate result relevance, and make decisions about which documents to download for further processing.
The search strategy involves multiple approaches to ensure comprehensive coverage of the topic. The agent starts with direct keyword searches but also employs semantic search techniques to find related concepts and alternative perspectives. It uses various search engines and academic databases to gather diverse sources of information.
The agent implements a scoring system to evaluate document relevance based on multiple factors including title relevance, content preview analysis, source credibility, and document type. This scoring helps prioritize which documents to download when dealing with large result sets.
Here's a detailed implementation of the web search component:
The following code example demonstrates how to implement a sophisticated web search agent that can discover and evaluate relevant documents. This implementation includes multiple search strategies and relevance scoring to ensure high-quality document selection.
import aiohttp
import asyncio
from bs4 import BeautifulSoup
from urllib.parse import urljoin, urlparse
import hashlib
from typing import List, Dict, Tuple
import re
class DocumentMetadata:
def __init__(self, url: str, title: str, content_type: str, relevance_score: float):
self.url = url
self.title = title
self.content_type = content_type
self.relevance_score = relevance_score
self.local_path = None
self.download_timestamp = None
class WebSearchAgent:
def __init__(self, config: Dict):
self.config = config
self.session = None
self.downloaded_documents = []
self.search_engines = {
'google': self.search_google,
'bing': self.search_bing,
'academic': self.search_academic_sources
}
async def search_and_download(self, topic: str, max_documents: int = 20) -> List[DocumentMetadata]:
"""Main method to search for and download relevant documents"""
async with aiohttp.ClientSession() as session:
self.session = session
# Generate search queries with different strategies
search_queries = self.generate_search_queries(topic)
# Search across multiple engines
all_results = []
for query in search_queries:
for engine_name, search_func in self.search_engines.items():
try:
results = await search_func(query)
all_results.extend(results)
except Exception as e:
print(f"Error searching with {engine_name}: {e}")
# Remove duplicates and score results
unique_results = self.deduplicate_results(all_results)
scored_results = self.score_relevance(unique_results, topic)
# Select top results for download
selected_results = sorted(scored_results,
key=lambda x: x.relevance_score,
reverse=True)[:max_documents]
# Download selected documents
downloaded_docs = []
for result in selected_results:
try:
doc = await self.download_document(result)
if doc:
downloaded_docs.append(doc)
except Exception as e:
print(f"Error downloading {result.url}: {e}")
return downloaded_docs
def generate_search_queries(self, topic: str) -> List[str]:
"""Generate multiple search query variations for comprehensive coverage"""
base_queries = [
topic,
f"{topic} overview",
f"{topic} introduction",
f"{topic} fundamentals",
f"{topic} guide",
f"{topic} tutorial",
f"{topic} research",
f"{topic} analysis"
]
# Add domain-specific variations
domain_variations = [
f"{topic} PDF",
f"{topic} whitepaper",
f"{topic} documentation",
f"{topic} case study"
]
return base_queries + domain_variations
async def search_google(self, query: str) -> List[DocumentMetadata]:
"""Search Google for relevant documents"""
# Note: In a real implementation, you would use Google Custom Search API
# This is a simplified example showing the structure
search_url = f"https://www.google.com/search?q={query}+filetype:pdf OR filetype:html"
try:
async with self.session.get(search_url, headers=self.get_headers()) as response:
html = await response.text()
soup = BeautifulSoup(html, 'html.parser')
results = []
for result_div in soup.find_all('div', class_='g'):
title_elem = result_div.find('h3')
link_elem = result_div.find('a')
if title_elem and link_elem:
title = title_elem.get_text()
url = link_elem.get('href')
if url and url.startswith('http'):
content_type = self.detect_content_type(url)
if content_type in ['pdf', 'html']:
doc = DocumentMetadata(url, title, content_type, 0.0)
results.append(doc)
return results
except Exception as e:
print(f"Google search error: {e}")
return []
def score_relevance(self, documents: List[DocumentMetadata], topic: str) -> List[DocumentMetadata]:
"""Score document relevance based on multiple factors"""
topic_keywords = topic.lower().split()
for doc in documents:
score = 0.0
title_lower = doc.title.lower()
# Title relevance scoring
for keyword in topic_keywords:
if keyword in title_lower:
score += 2.0
elif any(keyword in word for word in title_lower.split()):
score += 1.0
# Content type preference
if doc.content_type == 'pdf':
score += 1.5 # PDFs often contain more comprehensive content
elif doc.content_type == 'html':
score += 1.0
# URL structure analysis
url_lower = doc.url.lower()
if any(indicator in url_lower for indicator in ['research', 'academic', 'paper', 'study']):
score += 1.0
doc.relevance_score = score
return documents
This web search implementation demonstrates how to create a comprehensive document discovery system that can intelligently find and evaluate relevant content across the internet. The scoring system ensures that the most relevant documents are prioritized for download and processing.
DOCUMENT DOWNLOAD AND STORAGE MANAGEMENT
Once relevant documents are identified, the system must efficiently download and store them in a local folder structure. This component handles various file types, manages download failures, implements retry mechanisms, and organizes files in a logical directory structure that facilitates later processing.
The download manager must handle different content types gracefully, including PDF documents that may require special handling and HTML pages that need to be saved with their associated resources. It also implements intelligent naming schemes to avoid conflicts and ensure easy identification of downloaded content.
Storage management includes creating appropriate folder structures, handling file naming conflicts, and maintaining metadata about downloaded documents. The system also implements cleanup mechanisms to manage disk space and remove outdated or irrelevant files.
Here's a comprehensive implementation of the document download and storage system:
The following code example shows how to implement a robust document download and storage manager that handles various file types, implements retry mechanisms, and maintains organized file structures. This component ensures reliable document acquisition and storage for subsequent processing.
import aiofiles
import aiohttp
from pathlib import Path
import hashlib
import json
from datetime import datetime
import mimetypes
from urllib.parse import urlparse
import asyncio
class DocumentDownloader:
def __init__(self, base_storage_path: str = "./downloaded_documents"):
self.base_storage_path = Path(base_storage_path)
self.base_storage_path.mkdir(exist_ok=True)
self.metadata_file = self.base_storage_path / "download_metadata.json"
self.download_metadata = self.load_metadata()
async def download_document(self, doc_metadata: DocumentMetadata) -> Optional[DocumentMetadata]:
"""Download a single document with retry logic and proper storage"""
# Check if already downloaded
url_hash = self.generate_url_hash(doc_metadata.url)
if url_hash in self.download_metadata:
existing_path = Path(self.download_metadata[url_hash]['local_path'])
if existing_path.exists():
doc_metadata.local_path = str(existing_path)
return doc_metadata
# Create session-specific folder
session_folder = self.create_session_folder()
# Attempt download with retries
max_retries = 3
for attempt in range(max_retries):
try:
success = await self.attempt_download(doc_metadata, session_folder)
if success:
self.update_metadata(doc_metadata, url_hash)
return doc_metadata
except Exception as e:
print(f"Download attempt {attempt + 1} failed for {doc_metadata.url}: {e}")
if attempt == max_retries - 1:
print(f"Failed to download after {max_retries} attempts: {doc_metadata.url}")
else:
await asyncio.sleep(2 ** attempt) # Exponential backoff
return None
async def attempt_download(self, doc_metadata: DocumentMetadata, session_folder: Path) -> bool:
"""Attempt to download a single document"""
headers = {
'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36',
'Accept': 'text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8',
'Accept-Language': 'en-US,en;q=0.5',
'Accept-Encoding': 'gzip, deflate',
'Connection': 'keep-alive',
}
timeout = aiohttp.ClientTimeout(total=30)
async with aiohttp.ClientSession(timeout=timeout) as session:
async with session.get(doc_metadata.url, headers=headers) as response:
if response.status == 200:
content = await response.read()
# Determine file extension and name
file_extension = self.determine_file_extension(doc_metadata, response)
safe_filename = self.create_safe_filename(doc_metadata.title, file_extension)
file_path = session_folder / safe_filename
# Ensure unique filename
file_path = self.ensure_unique_filename(file_path)
# Write file
async with aiofiles.open(file_path, 'wb') as f:
await f.write(content)
# Update document metadata
doc_metadata.local_path = str(file_path)
doc_metadata.download_timestamp = datetime.now().isoformat()
# Validate download
if await self.validate_download(file_path, doc_metadata.content_type):
print(f"Successfully downloaded: {file_path}")
return True
else:
file_path.unlink() # Remove invalid file
return False
else:
print(f"HTTP {response.status} for {doc_metadata.url}")
return False
def create_session_folder(self) -> Path:
"""Create a timestamped folder for this download session"""
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
session_folder = self.base_storage_path / f"session_{timestamp}"
session_folder.mkdir(exist_ok=True)
return session_folder
def create_safe_filename(self, title: str, extension: str) -> str:
"""Create a safe filename from document title"""
# Remove or replace unsafe characters
safe_title = re.sub(r'[<>:"/\\|?*]', '_', title)
safe_title = re.sub(r'\s+', '_', safe_title)
safe_title = safe_title[:100] # Limit length
if not safe_title:
safe_title = "document"
return f"{safe_title}.{extension}"
def determine_file_extension(self, doc_metadata: DocumentMetadata, response) -> str:
"""Determine appropriate file extension"""
# First try content type from response headers
content_type = response.headers.get('content-type', '').lower()
if 'pdf' in content_type:
return 'pdf'
elif 'html' in content_type:
return 'html'
# Try URL extension
parsed_url = urlparse(doc_metadata.url)
path_extension = Path(parsed_url.path).suffix.lower()
if path_extension in ['.pdf', '.html', '.htm']:
return path_extension[1:] # Remove the dot
# Default based on detected content type
if doc_metadata.content_type == 'pdf':
return 'pdf'
else:
return 'html'
async def validate_download(self, file_path: Path, expected_type: str) -> bool:
"""Validate that downloaded file is of expected type"""
if not file_path.exists() or file_path.stat().st_size == 0:
return False
# Read first few bytes to check file signature
async with aiofiles.open(file_path, 'rb') as f:
header = await f.read(8)
if expected_type == 'pdf':
return header.startswith(b'%PDF')
elif expected_type == 'html':
# For HTML, check if it contains HTML tags
try:
async with aiofiles.open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
content = await f.read(1000)
return '<html' in content.lower() or '<!doctype html' in content.lower()
except:
return False
return True
def update_metadata(self, doc_metadata: DocumentMetadata, url_hash: str):
"""Update download metadata"""
self.download_metadata[url_hash] = {
'url': doc_metadata.url,
'title': doc_metadata.title,
'local_path': doc_metadata.local_path,
'download_timestamp': doc_metadata.download_timestamp,
'content_type': doc_metadata.content_type,
'relevance_score': doc_metadata.relevance_score
}
# Save metadata to file
with open(self.metadata_file, 'w') as f:
json.dump(self.download_metadata, f, indent=2)
This download and storage system provides robust handling of document acquisition with proper error handling, retry mechanisms, and organized file management. The metadata tracking ensures that the system can efficiently manage downloaded content and avoid unnecessary re-downloads.
CONTENT EXTRACTION FROM PDF AND HTML DOCUMENTS
After documents are successfully downloaded, the system must extract meaningful text content from various file formats. This component handles the complexity of parsing PDF documents and HTML pages, cleaning extracted text, and preparing content for further processing by the RAG system.
PDF extraction requires sophisticated handling of different PDF structures, including scanned documents that may require OCR processing, multi-column layouts, and documents with embedded images and tables. The system must preserve important structural information while extracting clean, readable text.
HTML processing involves parsing web page structures, removing navigation elements and advertisements, extracting main content areas, and handling various encoding issues. The processor must be intelligent enough to identify and extract the primary content while discarding irrelevant page elements.
Here's a comprehensive implementation of the content extraction system:
The following code example demonstrates how to implement a robust content extraction system that handles both PDF and HTML documents. This implementation includes advanced text cleaning, structure preservation, and quality validation to ensure high-quality content extraction for subsequent processing.
import PyPDF2
import pdfplumber
from bs4 import BeautifulSoup
import re
from typing import Dict, List, Optional
import aiofiles
from dataclasses import dataclass
import asyncio
from pathlib import Path
@dataclass
class ExtractedContent:
source_url: str
source_file: str
title: str
content_type: str
raw_text: str
cleaned_text: str
metadata: Dict
extraction_quality_score: float
class ContentProcessor:
def __init__(self, config: Dict):
self.config = config
self.min_content_length = config.get('min_content_length', 500)
self.max_content_length = config.get('max_content_length', 50000)
async def process_documents(self, documents: List[DocumentMetadata]) -> List[ExtractedContent]:
"""Process all downloaded documents and extract content"""
extracted_contents = []
for doc in documents:
if not doc.local_path or not Path(doc.local_path).exists():
continue
try:
if doc.content_type == 'pdf':
content = await self.extract_pdf_content(doc)
elif doc.content_type == 'html':
content = await self.extract_html_content(doc)
else:
continue
if content and self.validate_content_quality(content):
extracted_contents.append(content)
except Exception as e:
print(f"Error processing {doc.local_path}: {e}")
return extracted_contents
async def extract_pdf_content(self, doc: DocumentMetadata) -> Optional[ExtractedContent]:
"""Extract content from PDF documents using multiple methods"""
file_path = Path(doc.local_path)
# Try pdfplumber first (better for complex layouts)
try:
content = await self.extract_pdf_with_pdfplumber(file_path)
if content and len(content.strip()) > self.min_content_length:
return self.create_extracted_content(doc, content, 'pdfplumber')
except Exception as e:
print(f"pdfplumber extraction failed for {file_path}: {e}")
# Fallback to PyPDF2
try:
content = await self.extract_pdf_with_pypdf2(file_path)
if content and len(content.strip()) > self.min_content_length:
return self.create_extracted_content(doc, content, 'pypdf2')
except Exception as e:
print(f"PyPDF2 extraction failed for {file_path}: {e}")
return None
async def extract_pdf_with_pdfplumber(self, file_path: Path) -> str:
"""Extract PDF content using pdfplumber for better layout handling"""
extracted_text = []
with pdfplumber.open(file_path) as pdf:
for page_num, page in enumerate(pdf.pages):
try:
# Extract text with layout preservation
text = page.extract_text(layout=True)
if text:
# Clean up the text while preserving structure
cleaned_text = self.clean_pdf_text(text)
if cleaned_text.strip():
extracted_text.append(f"--- Page {page_num + 1} ---\n{cleaned_text}")
# Also extract tables if present
tables = page.extract_tables()
for table_num, table in enumerate(tables):
if table:
table_text = self.format_table_as_text(table)
extracted_text.append(f"--- Table {table_num + 1} on Page {page_num + 1} ---\n{table_text}")
except Exception as e:
print(f"Error extracting page {page_num + 1}: {e}")
continue
return "\n\n".join(extracted_text)
async def extract_pdf_with_pypdf2(self, file_path: Path) -> str:
"""Fallback PDF extraction using PyPDF2"""
extracted_text = []
with open(file_path, 'rb') as file:
pdf_reader = PyPDF2.PdfReader(file)
for page_num, page in enumerate(pdf_reader.pages):
try:
text = page.extract_text()
if text:
cleaned_text = self.clean_pdf_text(text)
if cleaned_text.strip():
extracted_text.append(f"--- Page {page_num + 1} ---\n{cleaned_text}")
except Exception as e:
print(f"Error extracting page {page_num + 1}: {e}")
continue
return "\n\n".join(extracted_text)
def clean_pdf_text(self, text: str) -> str:
"""Clean extracted PDF text while preserving important structure"""
# Remove excessive whitespace but preserve paragraph breaks
text = re.sub(r'\n\s*\n\s*\n+', '\n\n', text)
# Fix common PDF extraction issues
text = re.sub(r'([a-z])([A-Z])', r'\1 \2', text) # Add space between words
text = re.sub(r'(\w)-\s*\n\s*(\w)', r'\1\2', text) # Fix hyphenated words
text = re.sub(r'\s+', ' ', text) # Normalize whitespace
# Remove page headers/footers (common patterns)
lines = text.split('\n')
cleaned_lines = []
for line in lines:
line = line.strip()
# Skip likely headers/footers
if (len(line) < 10 or
re.match(r'^\d+$', line) or # Page numbers
re.match(r'^Page \d+', line) or
line.lower().startswith('copyright') or
line.count('.') > len(line) / 3): # Likely table of contents
continue
cleaned_lines.append(line)
return '\n'.join(cleaned_lines)
def format_table_as_text(self, table: List[List[str]]) -> str:
"""Convert extracted table to readable text format"""
if not table:
return ""
formatted_rows = []
for row in table:
if row and any(cell for cell in row if cell): # Skip empty rows
clean_row = [str(cell).strip() if cell else "" for cell in row]
formatted_rows.append(" | ".join(clean_row))
return "\n".join(formatted_rows)
async def extract_html_content(self, doc: DocumentMetadata) -> Optional[ExtractedContent]:
"""Extract content from HTML documents"""
file_path = Path(doc.local_path)
try:
async with aiofiles.open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
html_content = await f.read()
soup = BeautifulSoup(html_content, 'html.parser')
# Remove unwanted elements
self.remove_unwanted_elements(soup)
# Extract main content
main_content = self.extract_main_content(soup)
if main_content and len(main_content.strip()) > self.min_content_length:
return self.create_extracted_content(doc, main_content, 'beautifulsoup')
except Exception as e:
print(f"HTML extraction failed for {file_path}: {e}")
return None
def remove_unwanted_elements(self, soup: BeautifulSoup):
"""Remove navigation, ads, and other non-content elements"""
# Remove script and style elements
for element in soup(['script', 'style', 'nav', 'header', 'footer', 'aside']):
element.decompose()
# Remove elements with common non-content classes/ids
unwanted_selectors = [
'[class*="nav"]', '[class*="menu"]', '[class*="sidebar"]',
'[class*="ad"]', '[class*="advertisement"]', '[class*="banner"]',
'[id*="nav"]', '[id*="menu"]', '[id*="sidebar"]',
'[id*="ad"]', '[id*="advertisement"]', '[id*="banner"]'
]
for selector in unwanted_selectors:
for element in soup.select(selector):
element.decompose()
def extract_main_content(self, soup: BeautifulSoup) -> str:
"""Extract main content from cleaned HTML"""
# Try to find main content area
main_selectors = [
'main', 'article', '[role="main"]',
'.content', '.main-content', '.article-content',
'#content', '#main-content', '#article-content'
]
for selector in main_selectors:
main_element = soup.select_one(selector)
if main_element:
return self.clean_html_text(main_element.get_text())
# Fallback: extract from body
body = soup.find('body')
if body:
return self.clean_html_text(body.get_text())
# Last resort: entire document
return self.clean_html_text(soup.get_text())
def clean_html_text(self, text: str) -> str:
"""Clean extracted HTML text"""
# Normalize whitespace
text = re.sub(r'\s+', ' ', text)
text = re.sub(r'\n\s*\n\s*\n+', '\n\n', text)
# Remove excessive punctuation
text = re.sub(r'[.]{3,}', '...', text)
text = re.sub(r'[-]{3,}', '---', text)
return text.strip()
def create_extracted_content(self, doc: DocumentMetadata, text: str, extraction_method: str) -> ExtractedContent:
"""Create ExtractedContent object with quality scoring"""
cleaned_text = self.final_text_cleanup(text)
quality_score = self.calculate_quality_score(cleaned_text)
metadata = {
'extraction_method': extraction_method,
'original_length': len(text),
'cleaned_length': len(cleaned_text),
'source_type': doc.content_type
}
return ExtractedContent(
source_url=doc.url,
source_file=doc.local_path,
title=doc.title,
content_type=doc.content_type,
raw_text=text,
cleaned_text=cleaned_text,
metadata=metadata,
extraction_quality_score=quality_score
)
def final_text_cleanup(self, text: str) -> str:
"""Final cleanup of extracted text"""
# Remove very short lines (likely artifacts)
lines = text.split('\n')
cleaned_lines = [line for line in lines if len(line.strip()) > 10 or line.strip() == '']
# Rejoin and normalize
text = '\n'.join(cleaned_lines)
text = re.sub(r'\n{3,}', '\n\n', text)
return text.strip()
def calculate_quality_score(self, text: str) -> float:
"""Calculate content quality score based on various factors"""
if not text:
return 0.0
score = 0.0
# Length factor
length = len(text)
if length > self.min_content_length:
score += min(1.0, length / 5000) # Max 1.0 for 5000+ chars
# Sentence structure
sentences = re.split(r'[.!?]+', text)
avg_sentence_length = sum(len(s.split()) for s in sentences) / max(len(sentences), 1)
if 10 <= avg_sentence_length <= 30: # Good sentence length
score += 0.5
# Vocabulary diversity
words = re.findall(r'\b\w+\b', text.lower())
unique_words = set(words)
if words:
diversity = len(unique_words) / len(words)
score += min(0.5, diversity * 2) # Max 0.5 for high diversity
return min(1.0, score)
def validate_content_quality(self, content: ExtractedContent) -> bool:
"""Validate if extracted content meets quality standards"""
return (content.extraction_quality_score >= 0.3 and
len(content.cleaned_text) >= self.min_content_length and
len(content.cleaned_text) <= self.max_content_length)
This content extraction system provides comprehensive handling of both PDF and HTML documents with sophisticated text cleaning and quality validation. The multi-method approach ensures maximum content recovery while maintaining text quality for subsequent processing steps.
RAG SYSTEM IMPLEMENTATION
The Retrieval-Augmented Generation system forms the core intelligence of our presentation creation agent. This component indexes the extracted content, enables semantic search capabilities, and provides contextually relevant information retrieval for slide generation. The RAG system must efficiently handle large volumes of text while maintaining fast query response times.
The implementation involves creating vector embeddings of the extracted content, building efficient search indices, and implementing sophisticated retrieval strategies that can find relevant information based on semantic similarity rather than just keyword matching. The system also needs to handle content chunking to ensure optimal retrieval granularity.
The RAG system maintains context awareness across multiple documents and can synthesize information from various sources to provide comprehensive answers to queries about the presentation topic. It implements advanced ranking algorithms to ensure the most relevant content is prioritized in retrieval results.
Here's a detailed implementation of the RAG system:
The following code example demonstrates how to implement a sophisticated RAG system that can index extracted content, perform semantic search, and provide contextually relevant information retrieval. This implementation includes advanced chunking strategies, embedding generation, and retrieval optimization for presentation content generation.
import numpy as np
from sentence_transformers import SentenceTransformer
import faiss
from typing import List, Dict, Tuple, Optional
import pickle
import json
from dataclasses import dataclass, asdict
import re
from pathlib import Path
import asyncio
@dataclass
class ContentChunk:
chunk_id: str
source_url: str
source_file: str
content: str
chunk_index: int
total_chunks: int
embedding: Optional[np.ndarray] = None
metadata: Dict = None
@dataclass
class RetrievalResult:
chunk: ContentChunk
similarity_score: float
relevance_rank: int
class RAGSystem:
def __init__(self, config: Dict):
self.config = config
self.embedding_model_name = config.get('embedding_model', 'all-MiniLM-L6-v2')
self.chunk_size = config.get('chunk_size', 512)
self.chunk_overlap = config.get('chunk_overlap', 50)
self.max_chunks_per_query = config.get('max_chunks_per_query', 10)
# Initialize embedding model
self.embedding_model = SentenceTransformer(self.embedding_model_name)
self.embedding_dimension = self.embedding_model.get_sentence_embedding_dimension()
# Initialize FAISS index
self.index = faiss.IndexFlatIP(self.embedding_dimension) # Inner product for cosine similarity
self.chunks = []
self.chunk_id_to_index = {}
# Storage paths
self.storage_path = Path(config.get('rag_storage_path', './rag_storage'))
self.storage_path.mkdir(exist_ok=True)
async def index_content(self, extracted_contents: List[ExtractedContent]):
"""Index all extracted content for retrieval"""
print("Starting content indexing...")
# Clear existing index
self.index.reset()
self.chunks.clear()
self.chunk_id_to_index.clear()
all_chunks = []
# Process each document
for content in extracted_contents:
document_chunks = await self.create_chunks(content)
all_chunks.extend(document_chunks)
# Generate embeddings for all chunks
print(f"Generating embeddings for {len(all_chunks)} chunks...")
chunk_texts = [chunk.content for chunk in all_chunks]
embeddings = self.embedding_model.encode(chunk_texts, show_progress_bar=True)
# Normalize embeddings for cosine similarity
embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
# Add to FAISS index
self.index.add(embeddings.astype('float32'))
# Store chunks with embeddings
for i, (chunk, embedding) in enumerate(zip(all_chunks, embeddings)):
chunk.embedding = embedding
self.chunks.append(chunk)
self.chunk_id_to_index[chunk.chunk_id] = i
# Save index and chunks
await self.save_index()
print(f"Successfully indexed {len(all_chunks)} chunks from {len(extracted_contents)} documents")
async def create_chunks(self, content: ExtractedContent) -> List[ContentChunk]:
"""Create overlapping chunks from extracted content"""
text = content.cleaned_text
# Split into sentences for better chunk boundaries
sentences = self.split_into_sentences(text)
chunks = []
current_chunk = []
current_length = 0
chunk_index = 0
for sentence in sentences:
sentence_length = len(sentence.split())
# Check if adding this sentence would exceed chunk size
if current_length + sentence_length > self.chunk_size and current_chunk:
# Create chunk from current sentences
chunk_text = ' '.join(current_chunk)
chunk = self.create_chunk_object(content, chunk_text, chunk_index)
chunks.append(chunk)
# Start new chunk with overlap
overlap_sentences = self.get_overlap_sentences(current_chunk)
current_chunk = overlap_sentences + [sentence]
current_length = sum(len(s.split()) for s in current_chunk)
chunk_index += 1
else:
current_chunk.append(sentence)
current_length += sentence_length
# Add final chunk if it has content
if current_chunk:
chunk_text = ' '.join(current_chunk)
chunk = self.create_chunk_object(content, chunk_text, chunk_index)
chunks.append(chunk)
# Update total chunks count
for chunk in chunks:
chunk.total_chunks = len(chunks)
return chunks
def split_into_sentences(self, text: str) -> List[str]:
"""Split text into sentences using regex"""
# Split on sentence endings, but preserve the punctuation
sentences = re.split(r'(?<=[.!?])\s+', text)
# Filter out very short sentences and clean up
cleaned_sentences = []
for sentence in sentences:
sentence = sentence.strip()
if len(sentence) > 10: # Minimum sentence length
cleaned_sentences.append(sentence)
return cleaned_sentences
def get_overlap_sentences(self, sentences: List[str]) -> List[str]:
"""Get sentences for overlap between chunks"""
overlap_words = 0
overlap_sentences = []
# Take sentences from the end until we reach overlap word count
for sentence in reversed(sentences):
word_count = len(sentence.split())
if overlap_words + word_count <= self.chunk_overlap:
overlap_sentences.insert(0, sentence)
overlap_words += word_count
else:
break
return overlap_sentences
def create_chunk_object(self, content: ExtractedContent, chunk_text: str, chunk_index: int) -> ContentChunk:
"""Create a ContentChunk object"""
chunk_id = f"{hash(content.source_url)}_{chunk_index}"
metadata = {
'source_title': content.title,
'source_type': content.content_type,
'extraction_quality': content.extraction_quality_score,
'chunk_word_count': len(chunk_text.split())
}
return ContentChunk(
chunk_id=chunk_id,
source_url=content.source_url,
source_file=content.source_file,
content=chunk_text,
chunk_index=chunk_index,
total_chunks=0, # Will be updated later
metadata=metadata
)
async def retrieve_relevant_content(self, query: str, max_results: int = None) -> List[RetrievalResult]:
"""Retrieve relevant content chunks for a given query"""
if max_results is None:
max_results = self.max_chunks_per_query
if not self.chunks:
print("No content indexed. Please index content first.")
return []
# Generate query embedding
query_embedding = self.embedding_model.encode([query])
query_embedding = query_embedding / np.linalg.norm(query_embedding, axis=1, keepdims=True)
# Search in FAISS index
similarities, indices = self.index.search(query_embedding.astype('float32'), max_results)
# Create retrieval results
results = []
for rank, (similarity, index) in enumerate(zip(similarities[0], indices[0])):
if index < len(self.chunks): # Valid index
chunk = self.chunks[index]
result = RetrievalResult(
chunk=chunk,
similarity_score=float(similarity),
relevance_rank=rank
)
results.append(result)
# Apply additional ranking based on content quality and diversity
results = self.rerank_results(results, query)
return results
def rerank_results(self, results: List[RetrievalResult], query: str) -> List[RetrievalResult]:
"""Apply additional ranking to improve result quality and diversity"""
# Calculate additional relevance factors
for result in results:
chunk = result.chunk
# Keyword overlap bonus
query_words = set(query.lower().split())
chunk_words = set(chunk.content.lower().split())
keyword_overlap = len(query_words.intersection(chunk_words)) / len(query_words)
# Content quality bonus
quality_bonus = chunk.metadata.get('extraction_quality', 0.5)
# Combine scores
combined_score = (result.similarity_score * 0.7 +
keyword_overlap * 0.2 +
quality_bonus * 0.1)
result.similarity_score = combined_score
# Re-sort by combined score
results.sort(key=lambda x: x.similarity_score, reverse=True)
# Update ranks
for i, result in enumerate(results):
result.relevance_rank = i
# Apply diversity filtering to avoid too many chunks from same source
diverse_results = self.apply_diversity_filter(results)
return diverse_results
def apply_diversity_filter(self, results: List[RetrievalResult], max_per_source: int = 3) -> List[RetrievalResult]:
"""Filter results to ensure diversity across sources"""
source_counts = {}
filtered_results = []
for result in results:
source_url = result.chunk.source_url
current_count = source_counts.get(source_url, 0)
if current_count < max_per_source:
filtered_results.append(result)
source_counts[source_url] = current_count + 1
return filtered_results
async def generate_context_for_slide(self, slide_topic: str, slide_context: str = "") -> str:
"""Generate contextual information for a specific slide topic"""
# Combine slide topic with any additional context
query = f"{slide_topic} {slide_context}".strip()
# Retrieve relevant content
results = await self.retrieve_relevant_content(query, max_results=5)
if not results:
return f"No relevant content found for: {slide_topic}"
# Synthesize content from top results
context_parts = []
for result in results:
chunk = result.chunk
context_parts.append(f"From {chunk.metadata['source_title']}:\n{chunk.content}")
synthesized_context = "\n\n---\n\n".join(context_parts)
return synthesized_context
async def save_index(self):
"""Save FAISS index and chunks to disk"""
# Save FAISS index
index_path = self.storage_path / "faiss_index.bin"
faiss.write_index(self.index, str(index_path))
# Save chunks (without embeddings to save space)
chunks_data = []
for chunk in self.chunks:
chunk_dict = asdict(chunk)
chunk_dict['embedding'] = None # Don't save embeddings separately
chunks_data.append(chunk_dict)
chunks_path = self.storage_path / "chunks.json"
with open(chunks_path, 'w') as f:
json.dump(chunks_data, f, indent=2)
# Save metadata
metadata = {
'embedding_model': self.embedding_model_name,
'embedding_dimension': self.embedding_dimension,
'chunk_count': len(self.chunks),
'chunk_size': self.chunk_size,
'chunk_overlap': self.chunk_overlap
}
metadata_path = self.storage_path / "index_metadata.json"
with open(metadata_path, 'w') as f:
json.dump(metadata, f, indent=2)
print(f"Index saved to {self.storage_path}")
async def load_index(self) -> bool:
"""Load FAISS index and chunks from disk"""
try:
# Load FAISS index
index_path = self.storage_path / "faiss_index.bin"
if not index_path.exists():
return False
self.index = faiss.read_index(str(index_path))
# Load chunks
chunks_path = self.storage_path / "chunks.json"
with open(chunks_path, 'r') as f:
chunks_data = json.load(f)
self.chunks = []
self.chunk_id_to_index = {}
for i, chunk_dict in enumerate(chunks_data):
chunk = ContentChunk(**chunk_dict)
self.chunks.append(chunk)
self.chunk_id_to_index[chunk.chunk_id] = i
print(f"Loaded index with {len(self.chunks)} chunks")
return True
except Exception as e:
print(f"Error loading index: {e}")
return False
This RAG system implementation provides sophisticated content indexing and retrieval capabilities with semantic search, intelligent chunking, and advanced ranking algorithms. The system ensures that the most relevant and high-quality content is available for presentation generation while maintaining fast query performance.
GRAPHRAG AND ONTOLOGY CREATION
The GraphRAG component extends traditional RAG capabilities by creating knowledge graphs and ontologies from the extracted content. This system identifies entities, relationships, and concepts within the documents, building a structured representation of knowledge that enables more sophisticated reasoning and content organization for presentation creation.
The ontology creation process involves named entity recognition, relationship extraction, and concept clustering to build a comprehensive knowledge graph. This structured representation allows the system to understand connections between different concepts and generate more coherent and logically organized presentations.
The GraphRAG system also implements graph-based retrieval algorithms that can traverse relationships to find related concepts and supporting information. This capability enables the creation of presentations with better narrative flow and logical progression between slides.
Here's a comprehensive implementation of the GraphRAG and ontology system:
The following code example demonstrates how to implement a sophisticated GraphRAG system that creates knowledge graphs and ontologies from extracted content. This implementation includes entity recognition, relationship extraction, and graph-based retrieval for enhanced presentation generation capabilities.
import spacy
import networkx as nx
from collections import defaultdict, Counter
import json
from typing import Dict, List, Tuple, Set, Optional
from dataclasses import dataclass, asdict
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.cluster import KMeans
from sklearn.metrics.pairwise import cosine_similarity
import pickle
@dataclass
class Entity:
text: str
label: str
confidence: float
mentions: List[str]
context_chunks: List[str]
@dataclass
class Relationship:
source_entity: str
target_entity: str
relation_type: str
confidence: float
evidence_text: str
source_chunk: str
@dataclass
class Concept:
name: str
keywords: List[str]
related_entities: List[str]
importance_score: float
chunk_ids: List[str]
class GraphRAGSystem:
def __init__(self, config: Dict):
self.config = config
self.nlp = spacy.load("en_core_web_sm")
# Knowledge graph
self.knowledge_graph = nx.DiGraph()
self.entities = {}
self.relationships = []
self.concepts = {}
# Configuration parameters
self.min_entity_confidence = config.get('min_entity_confidence', 0.7)
self.max_entities_per_chunk = config.get('max_entities_per_chunk', 20)
self.concept_cluster_count = config.get('concept_cluster_count', 10)
# Storage
self.storage_path = Path(config.get('graphrag_storage_path', './graphrag_storage'))
self.storage_path.mkdir(exist_ok=True)
async def create_ontology(self, extracted_contents: List[ExtractedContent]) -> Dict:
"""Create comprehensive ontology from extracted content"""
print("Creating ontology from extracted content...")
# Step 1: Extract entities from all content
all_chunks = []
for content in extracted_contents:
chunks = await self.extract_entities_from_content(content)
all_chunks.extend(chunks)
# Step 2: Extract relationships between entities
await self.extract_relationships(all_chunks)
# Step 3: Identify key concepts and themes
await self.identify_concepts(all_chunks)
# Step 4: Build knowledge graph
await self.build_knowledge_graph()
# Step 5: Calculate entity and concept importance
await self.calculate_importance_scores()
# Step 6: Create ontology structure
ontology = await self.create_ontology_structure()
# Step 7: Save ontology
await self.save_ontology(ontology)
print(f"Ontology created with {len(self.entities)} entities, {len(self.relationships)} relationships, and {len(self.concepts)} concepts")
return ontology
async def extract_entities_from_content(self, content: ExtractedContent) -> List[Dict]:
"""Extract named entities from content chunks"""
# Split content into chunks for processing
chunks = self.split_content_for_ner(content.cleaned_text)
processed_chunks = []
for i, chunk_text in enumerate(chunks):
chunk_id = f"{hash(content.source_url)}_{i}"
# Process with spaCy
doc = self.nlp(chunk_text)
chunk_entities = []
for ent in doc.ents:
# Filter entities by confidence and relevance
if (len(ent.text.strip()) > 2 and
ent.label_ in ['PERSON', 'ORG', 'GPE', 'PRODUCT', 'EVENT', 'WORK_OF_ART', 'LAW', 'LANGUAGE'] and
self.is_valid_entity(ent.text)):
entity_data = {
'text': ent.text.strip(),
'label': ent.label_,
'start': ent.start_char,
'end': ent.end_char,
'confidence': self.calculate_entity_confidence(ent, doc)
}
chunk_entities.append(entity_data)
# Limit entities per chunk to avoid noise
chunk_entities = sorted(chunk_entities, key=lambda x: x['confidence'], reverse=True)[:self.max_entities_per_chunk]
processed_chunk = {
'chunk_id': chunk_id,
'text': chunk_text,
'entities': chunk_entities,
'source_url': content.source_url,
'source_title': content.title
}
processed_chunks.append(processed_chunk)
# Update global entities dictionary
for entity_data in chunk_entities:
self.update_global_entity(entity_data, chunk_id, chunk_text)
return processed_chunks
def split_content_for_ner(self, text: str, max_chunk_size: int = 1000000) -> List[str]:
"""Split content into chunks suitable for NER processing"""
# spaCy has limits on text length, so we split long texts
if len(text) <= max_chunk_size:
return [text]
# Split on paragraph boundaries
paragraphs = text.split('\n\n')
chunks = []
current_chunk = []
current_length = 0
for paragraph in paragraphs:
para_length = len(paragraph)
if current_length + para_length > max_chunk_size and current_chunk:
chunks.append('\n\n'.join(current_chunk))
current_chunk = [paragraph]
current_length = para_length
else:
current_chunk.append(paragraph)
current_length += para_length
if current_chunk:
chunks.append('\n\n'.join(current_chunk))
return chunks
def is_valid_entity(self, entity_text: str) -> bool:
"""Check if entity text is valid and meaningful"""
entity_text = entity_text.strip()
# Filter out common false positives
if (len(entity_text) < 3 or
entity_text.isdigit() or
entity_text.lower() in ['the', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for', 'of', 'with', 'by'] or
entity_text.count(' ') > 5): # Very long entities are often errors
return False
return True
def calculate_entity_confidence(self, ent, doc) -> float:
"""Calculate confidence score for an entity"""
base_confidence = 0.8 # Base confidence for spaCy entities
# Adjust based on entity characteristics
if ent.text.istitle(): # Proper capitalization
base_confidence += 0.1
if len(ent.text.split()) > 1: # Multi-word entities often more reliable
base_confidence += 0.05
# Check if entity appears multiple times (higher confidence)
entity_count = sum(1 for token in doc if token.text.lower() == ent.text.lower())
if entity_count > 1:
base_confidence += min(0.1, entity_count * 0.02)
return min(1.0, base_confidence)
def update_global_entity(self, entity_data: Dict, chunk_id: str, chunk_text: str):
"""Update global entities dictionary with new entity occurrence"""
entity_text = entity_data['text'].lower()
if entity_text not in self.entities:
self.entities[entity_text] = Entity(
text=entity_data['text'],
label=entity_data['label'],
confidence=entity_data['confidence'],
mentions=[entity_data['text']],
context_chunks=[chunk_id]
)
else:
# Update existing entity
existing_entity = self.entities[entity_text]
existing_entity.mentions.append(entity_data['text'])
existing_entity.context_chunks.append(chunk_id)
# Update confidence (weighted average)
total_mentions = len(existing_entity.mentions)
existing_entity.confidence = ((existing_entity.confidence * (total_mentions - 1) +
entity_data['confidence']) / total_mentions)
async def extract_relationships(self, processed_chunks: List[Dict]):
"""Extract relationships between entities"""
print("Extracting relationships between entities...")
for chunk in processed_chunks:
chunk_text = chunk['text']
entities = chunk['entities']
# Extract relationships within each chunk
for i, entity1 in enumerate(entities):
for j, entity2 in enumerate(entities[i+1:], i+1):
relationship = self.identify_relationship(entity1, entity2, chunk_text)
if relationship:
relationship.source_chunk = chunk['chunk_id']
self.relationships.append(relationship)
def identify_relationship(self, entity1: Dict, entity2: Dict, text: str) -> Optional[Relationship]:
"""Identify relationship between two entities in text"""
# Simple pattern-based relationship extraction
e1_text = entity1['text']
e2_text = entity2['text']
# Find sentences containing both entities
sentences = text.split('.')
for sentence in sentences:
if e1_text in sentence and e2_text in sentence:
# Look for relationship patterns
sentence_lower = sentence.lower()
relationship_patterns = {
'is_part_of': ['part of', 'component of', 'element of', 'belongs to'],
'related_to': ['related to', 'associated with', 'connected to', 'linked to'],
'used_by': ['used by', 'utilized by', 'employed by'],
'created_by': ['created by', 'developed by', 'made by', 'built by'],
'located_in': ['located in', 'situated in', 'found in', 'based in'],
'works_for': ['works for', 'employed by', 'member of']
}
for relation_type, patterns in relationship_patterns.items():
for pattern in patterns:
if pattern in sentence_lower:
# Determine direction based on entity positions
e1_pos = sentence.find(e1_text)
e2_pos = sentence.find(e2_text)
pattern_pos = sentence_lower.find(pattern)
if e1_pos < pattern_pos < e2_pos:
source, target = e1_text, e2_text
elif e2_pos < pattern_pos < e1_pos:
source, target = e2_text, e1_text
else:
continue
return Relationship(
source_entity=source,
target_entity=target,
relation_type=relation_type,
confidence=0.7,
evidence_text=sentence.strip(),
source_chunk=""
)
# If no explicit relationship found, create a general co-occurrence relationship
if abs(entity1['start'] - entity2['start']) < 200: # Entities are close in text
return Relationship(
source_entity=e1_text,
target_entity=e2_text,
relation_type='co_occurs_with',
confidence=0.4,
evidence_text=text[max(0, min(entity1['start'], entity2['start'])-50):
max(entity1['end'], entity2['end'])+50],
source_chunk=""
)
return None
async def identify_concepts(self, processed_chunks: List[Dict]):
"""Identify key concepts and themes using clustering"""
print("Identifying key concepts and themes...")
# Prepare text data for clustering
chunk_texts = [chunk['text'] for chunk in processed_chunks]
# Use TF-IDF to vectorize text
vectorizer = TfidfVectorizer(
max_features=1000,
stop_words='english',
ngram_range=(1, 3),
min_df=2,
max_df=0.8
)
tfidf_matrix = vectorizer.fit_transform(chunk_texts)
feature_names = vectorizer.get_feature_names_out()
# Perform clustering
n_clusters = min(self.concept_cluster_count, len(chunk_texts))
if n_clusters > 1:
kmeans = KMeans(n_clusters=n_clusters, random_state=42)
cluster_labels = kmeans.fit_predict(tfidf_matrix)
# Extract concepts from clusters
for cluster_id in range(n_clusters):
cluster_chunks = [chunk for i, chunk in enumerate(processed_chunks) if cluster_labels[i] == cluster_id]
if cluster_chunks:
concept = self.create_concept_from_cluster(cluster_id, cluster_chunks, tfidf_matrix, feature_names, cluster_labels)
self.concepts[concept.name] = concept
def create_concept_from_cluster(self, cluster_id: int, cluster_chunks: List[Dict],
tfidf_matrix, feature_names: List[str], cluster_labels: np.ndarray) -> Concept:
"""Create a concept from a cluster of chunks"""
# Get cluster indices
cluster_indices = [i for i, label in enumerate(cluster_labels) if label == cluster_id]
# Calculate cluster centroid
cluster_tfidf = tfidf_matrix[cluster_indices]
centroid = np.mean(cluster_tfidf.toarray(), axis=0)
# Get top keywords for this cluster
top_indices = np.argsort(centroid)[-10:][::-1]
top_keywords = [feature_names[i] for i in top_indices if centroid[i] > 0]
# Extract entities from cluster chunks
cluster_entities = set()
chunk_ids = []
for chunk in cluster_chunks:
chunk_ids.append(chunk['chunk_id'])
for entity in chunk['entities']:
cluster_entities.add(entity['text'])
# Create concept name from top keywords
concept_name = f"concept_{cluster_id}_{top_keywords[0] if top_keywords else 'unknown'}"
# Calculate importance score based on cluster size and keyword strength
importance_score = len(cluster_chunks) / len(cluster_labels) * np.max(centroid)
return Concept(
name=concept_name,
keywords=top_keywords,
related_entities=list(cluster_entities),
importance_score=importance_score,
chunk_ids=chunk_ids
)
async def build_knowledge_graph(self):
"""Build NetworkX knowledge graph from entities and relationships"""
print("Building knowledge graph...")
# Add entity nodes
for entity_text, entity in self.entities.items():
self.knowledge_graph.add_node(
entity_text,
label=entity.label,
confidence=entity.confidence,
mention_count=len(entity.mentions),
type='entity'
)
# Add concept nodes
for concept_name, concept in self.concepts.items():
self.knowledge_graph.add_node(
concept_name,
keywords=concept.keywords,
importance=concept.importance_score,
type='concept'
)
# Connect concepts to related entities
for entity_text in concept.related_entities:
if entity_text.lower() in self.entities:
self.knowledge_graph.add_edge(
concept_name,
entity_text.lower(),
relation_type='contains_entity',
weight=0.5
)
# Add relationship edges
for relationship in self.relationships:
source = relationship.source_entity.lower()
target = relationship.target_entity.lower()
if source in self.entities and target in self.entities:
self.knowledge_graph.add_edge(
source,
target,
relation_type=relationship.relation_type,
confidence=relationship.confidence,
evidence=relationship.evidence_text
)
async def calculate_importance_scores(self):
"""Calculate importance scores using graph centrality measures"""
print("Calculating importance scores...")
if len(self.knowledge_graph.nodes()) > 0:
# Calculate various centrality measures
pagerank_scores = nx.pagerank(self.knowledge_graph)
betweenness_scores = nx.betweenness_centrality(self.knowledge_graph)
degree_scores = dict(self.knowledge_graph.degree())
# Update entity importance scores
for entity_text, entity in self.entities.items():
if entity_text in pagerank_scores:
# Combine different centrality measures
combined_score = (pagerank_scores[entity_text] * 0.4 +
betweenness_scores[entity_text] * 0.3 +
degree_scores[entity_text] / max(degree_scores.values()) * 0.3)
# Also factor in mention frequency
mention_bonus = min(0.3, len(entity.mentions) / 10)
entity.confidence = min(1.0, entity.confidence + combined_score + mention_bonus)
async def create_ontology_structure(self) -> Dict:
"""Create structured ontology representation"""
# Sort entities by importance
sorted_entities = sorted(self.entities.items(),
key=lambda x: x[1].confidence,
reverse=True)
# Sort concepts by importance
sorted_concepts = sorted(self.concepts.items(),
key=lambda x: x[1].importance_score,
reverse=True)
ontology = {
'entities': {
entity_text: asdict(entity) for entity_text, entity in sorted_entities[:50] # Top 50 entities
},
'concepts': {
concept_name: asdict(concept) for concept_name, concept in sorted_concepts[:20] # Top 20 concepts
},
'relationships': [asdict(rel) for rel in self.relationships if rel.confidence > 0.5],
'graph_statistics': {
'total_nodes': len(self.knowledge_graph.nodes()),
'total_edges': len(self.knowledge_graph.edges()),
'average_degree': sum(dict(self.knowledge_graph.degree()).values()) / len(self.knowledge_graph.nodes()) if len(self.knowledge_graph.nodes()) > 0 else 0
}
}
return ontology
async def get_related_concepts(self, entity_or_concept: str, max_results: int = 5) -> List[Tuple[str, float]]:
"""Get concepts related to a given entity or concept"""
if entity_or_concept.lower() not in self.knowledge_graph:
return []
# Use graph traversal to find related nodes
related_nodes = []
# Direct neighbors
for neighbor in self.knowledge_graph.neighbors(entity_or_concept.lower()):
edge_data = self.knowledge_graph.get_edge_data(entity_or_concept.lower(), neighbor)
weight = edge_data.get('confidence', 0.5) if edge_data else 0.5
related_nodes.append((neighbor, weight))
# Sort by weight and return top results
related_nodes.sort(key=lambda x: x[1], reverse=True)
return related_nodes[:max_results]
async def save_ontology(self, ontology: Dict):
"""Save ontology and knowledge graph to disk"""
# Save ontology JSON
ontology_path = self.storage_path / "ontology.json"
with open(ontology_path, 'w') as f:
json.dump(ontology, f, indent=2)
# Save knowledge graph
graph_path = self.storage_path / "knowledge_graph.pickle"
with open(graph_path, 'wb') as f:
pickle.dump(self.knowledge_graph, f)
print(f"Ontology saved to {self.storage_path}")
This GraphRAG implementation creates sophisticated knowledge representations that enable the system to understand relationships between concepts and generate more coherent, well-structured presentations. The ontology provides a foundation for intelligent content organization and narrative flow in the generated slides.
VISUAL CONTENT EXTRACTION WITH VLM
The Visual Language Model component enhances the presentation creation system by extracting and analyzing visual elements from the downloaded documents. This component identifies charts, diagrams, images, and figures that can enhance the presentation's visual appeal and information delivery effectiveness.
The VLM system processes both PDF and HTML documents to locate visual content, extracts relevant images and charts, analyzes their content to understand their purpose and relevance, and prepares them for inclusion in the generated presentations. The system also generates appropriate captions and descriptions for accessibility and context.
The visual processor implements intelligent filtering to ensure only high-quality, relevant visual content is selected for inclusion in presentations. It also handles image format conversion and sizing optimization to ensure compatibility with PowerPoint requirements.
Here's a detailed implementation of the visual content extraction system:
The following code example shows how to implement a sophisticated visual content extraction system using Vision Language Models to identify, extract, and analyze visual elements from documents. This implementation includes image quality assessment, relevance scoring, and content analysis for intelligent visual content selection.
import cv2
import numpy as np
from PIL import Image, ImageEnhance
import fitz # PyMuPDF
from bs4 import BeautifulSoup
import requests
from urllib.parse import urljoin, urlparse
import base64
import io
from typing import List, Dict, Tuple, Optional
from dataclasses import dataclass
import asyncio
import aiohttp
import aiofiles
from pathlib import Path
import hashlib
@dataclass
class VisualElement:
element_id: str
source_url: str
source_file: str
element_type: str # 'image', 'chart', 'diagram', 'table'
local_path: str
original_caption: str
generated_caption: str
relevance_score: float
quality_score: float
dimensions: Tuple[int, int]
file_size: int
format: str
page_number: Optional[int] = None
extraction_method: str = ""
class VisualProcessor:
def __init__(self, config: Dict):
self.config = config
self.min_image_size = config.get('min_image_size', (100, 100))
self.max_image_size = config.get('max_image_size', (2000, 2000))
self.min_quality_score = config.get('min_quality_score', 0.5)
self.supported_formats = config.get('supported_formats', ['png', 'jpg', 'jpeg', 'gif', 'bmp'])
# Storage setup
self.visual_storage_path = Path(config.get('visual_storage_path', './visual_content'))
self.visual_storage_path.mkdir(exist_ok=True)
# VLM setup (placeholder for actual VLM integration)
self.vlm_enabled = config.get('vlm_enabled', False)
async def extract_visuals(self, documents: List[DocumentMetadata]) -> List[VisualElement]:
"""Extract visual elements from all documents"""
print("Extracting visual elements from documents...")
all_visuals = []
for doc in documents:
if not doc.local_path or not Path(doc.local_path).exists():
continue
try:
if doc.content_type == 'pdf':
visuals = await self.extract_pdf_visuals(doc)
elif doc.content_type == 'html':
visuals = await self.extract_html_visuals(doc)
else:
continue
# Filter and validate visuals
validated_visuals = await self.validate_and_filter_visuals(visuals)
all_visuals.extend(validated_visuals)
except Exception as e:
print(f"Error extracting visuals from {doc.local_path}: {e}")
# Sort by relevance and quality
all_visuals.sort(key=lambda x: (x.relevance_score + x.quality_score) / 2, reverse=True)
print(f"Extracted {len(all_visuals)} visual elements")
return all_visuals
async def extract_pdf_visuals(self, doc: DocumentMetadata) -> List[VisualElement]:
"""Extract visual elements from PDF documents"""
visuals = []
pdf_path = Path(doc.local_path)
try:
# Open PDF with PyMuPDF
pdf_document = fitz.open(pdf_path)
for page_num in range(len(pdf_document)):
page = pdf_document.load_page(page_num)
# Extract images from page
image_list = page.get_images()
for img_index, img in enumerate(image_list):
try:
# Get image data
xref = img[0]
pix = fitz.Pixmap(pdf_document, xref)
# Skip if image is too small or in unsupported format
if pix.width < self.min_image_size[0] or pix.height < self.min_image_size[1]:
pix = None
continue
# Convert to PIL Image
if pix.n - pix.alpha < 4: # GRAY or RGB
img_data = pix.tobytes("png")
img_pil = Image.open(io.BytesIO(img_data))
else: # CMYK
pix1 = fitz.Pixmap(fitz.csRGB, pix)
img_data = pix1.tobytes("png")
img_pil = Image.open(io.BytesIO(img_data))
pix1 = None
pix = None
# Save image and create visual element
visual_element = await self.create_visual_element_from_image(
img_pil, doc, page_num, img_index, 'pdf_extraction'
)
if visual_element:
visuals.append(visual_element)
except Exception as e:
print(f"Error extracting image {img_index} from page {page_num}: {e}")
continue
# Extract vector graphics and charts (simplified approach)
# This would require more sophisticated analysis in a real implementation
drawings = page.get_drawings()
if drawings:
# Create a rendered image of the page for chart detection
mat = fitz.Matrix(2, 2) # 2x zoom
pix = page.get_pixmap(matrix=mat)
img_data = pix.tobytes("png")
page_image = Image.open(io.BytesIO(img_data))
# Analyze for chart-like content
if await self.detect_chart_content(page_image):
visual_element = await self.create_visual_element_from_image(
page_image, doc, page_num, 0, 'chart_detection'
)
if visual_element:
visual_element.element_type = 'chart'
visuals.append(visual_element)
pdf_document.close()
except Exception as e:
print(f"Error processing PDF {pdf_path}: {e}")
return visuals
async def extract_html_visuals(self, doc: DocumentMetadata) -> List[VisualElement]:
"""Extract visual elements from HTML documents"""
visuals = []
html_path = Path(doc.local_path)
try:
async with aiofiles.open(html_path, 'r', encoding='utf-8', errors='ignore') as f:
html_content = await f.read()
soup = BeautifulSoup(html_content, 'html.parser')
# Find all image elements
img_tags = soup.find_all('img')
for img_index, img_tag in enumerate(img_tags):
try:
img_src = img_tag.get('src')
if not img_src:
continue
# Handle relative URLs
if not img_src.startswith('http'):
img_src = urljoin(doc.url, img_src)
# Download and process image
img_pil = await self.download_image_from_url(img_src)
if img_pil:
# Get caption from alt text or nearby text
caption = img_tag.get('alt', '') or img_tag.get('title', '')
if not caption:
caption = self.extract_nearby_text(img_tag)
visual_element = await self.create_visual_element_from_image(
img_pil, doc, None, img_index, 'html_extraction'
)
if visual_element:
visual_element.original_caption = caption
visuals.append(visual_element)
except Exception as e:
print(f"Error processing HTML image {img_index}: {e}")
continue
# Look for SVG elements (charts/diagrams)
svg_tags = soup.find_all('svg')
for svg_index, svg_tag in enumerate(svg_tags):
try:
# Convert SVG to image (simplified approach)
svg_content = str(svg_tag)
if len(svg_content) > 100: # Skip very small SVGs
# In a real implementation, you would use a library like cairosvg
# to convert SVG to PNG
pass
except Exception as e:
print(f"Error processing SVG {svg_index}: {e}")
continue
except Exception as e:
print(f"Error processing HTML {html_path}: {e}")
return visuals
async def download_image_from_url(self, url: str) -> Optional[Image.Image]:
"""Download image from URL and return PIL Image"""
try:
timeout = aiohttp.ClientTimeout(total=10)
headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'}
async with aiohttp.ClientSession(timeout=timeout) as session:
async with session.get(url, headers=headers) as response:
if response.status == 200:
image_data = await response.read()
return Image.open(io.BytesIO(image_data))
except Exception as e:
print(f"Error downloading image from {url}: {e}")
return None
def extract_nearby_text(self, img_tag) -> str:
"""Extract text near an image tag for caption generation"""
# Look for captions in nearby elements
parent = img_tag.parent
if parent:
# Check for figure caption
figcaption = parent.find('figcaption')
if figcaption:
return figcaption.get_text().strip()
# Check for nearby text
siblings = parent.find_all(text=True)
nearby_text = ' '.join([text.strip() for text in siblings if text.strip()])
if len(nearby_text) > 10:
return nearby_text[:200] # Limit length
return ""
async def create_visual_element_from_image(self, img_pil: Image.Image, doc: DocumentMetadata,
page_num: Optional[int], img_index: int,
extraction_method: str) -> Optional[VisualElement]:
"""Create VisualElement from PIL Image"""
try:
# Generate unique ID
element_id = hashlib.md5(f"{doc.url}_{page_num}_{img_index}".encode()).hexdigest()
# Check image dimensions and quality
width, height = img_pil.size
if width < self.min_image_size[0] or height < self.min_image_size[1]:
return None
# Calculate quality score
quality_score = await self.calculate_image_quality(img_pil)
if quality_score < self.min_quality_score:
return None
# Optimize image for presentation use
optimized_img = await self.optimize_image_for_presentation(img_pil)
# Save image
filename = f"{element_id}.png"
local_path = self.visual_storage_path / filename
optimized_img.save(local_path, 'PNG', optimize=True)
# Determine element type
element_type = await self.classify_visual_element(optimized_img)
# Calculate relevance score (placeholder - would use VLM in real implementation)
relevance_score = await self.calculate_relevance_score(optimized_img, doc.title)
# Generate caption using VLM (placeholder)
generated_caption = await self.generate_image_caption(optimized_img)
return VisualElement(
element_id=element_id,
source_url=doc.url,
source_file=doc.local_path,
element_type=element_type,
local_path=str(local_path),
original_caption="",
generated_caption=generated_caption,
relevance_score=relevance_score,
quality_score=quality_score,
dimensions=(optimized_img.width, optimized_img.height),
file_size=local_path.stat().st_size,
format='png',
page_number=page_num,
extraction_method=extraction_method
)
except Exception as e:
print(f"Error creating visual element: {e}")
return None
async def calculate_image_quality(self, img: Image.Image) -> float:
"""Calculate image quality score based on various factors"""
# Convert to numpy array for analysis
img_array = np.array(img.convert('RGB'))
quality_score = 0.0
# Check resolution
width, height = img.size
pixel_count = width * height
resolution_score = min(1.0, pixel_count / (500 * 500)) # Normalize to 500x500
quality_score += resolution_score * 0.3
# Check for blur (using Laplacian variance)
gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
blur_score = cv2.Laplacian(gray, cv2.CV_64F).var()
normalized_blur = min(1.0, blur_score / 1000) # Normalize
quality_score += normalized_blur * 0.3
# Check contrast
contrast = np.std(gray)
normalized_contrast = min(1.0, contrast / 64) # Normalize
quality_score += normalized_contrast * 0.2
# Check for mostly white/empty images
mean_brightness = np.mean(gray)
if mean_brightness > 240: # Very bright (likely empty)
quality_score *= 0.5
# Check aspect ratio (prefer reasonable ratios)
aspect_ratio = max(width, height) / min(width, height)
if aspect_ratio > 5: # Very wide or tall images
quality_score *= 0.7
quality_score += 0.2 # Base score
return min(1.0, quality_score)
async def optimize_image_for_presentation(self, img: Image.Image) -> Image.Image:
"""Optimize image for presentation use"""
# Convert to RGB if necessary
if img.mode != 'RGB':
img = img.convert('RGB')
# Resize if too large
width, height = img.size
if width > self.max_image_size[0] or height > self.max_image_size[1]:
img.thumbnail(self.max_image_size, Image.Resampling.LANCZOS)
# Enhance contrast and sharpness slightly
enhancer = ImageEnhance.Contrast(img)
img = enhancer.enhance(1.1)
enhancer = ImageEnhance.Sharpness(img)
img = enhancer.enhance(1.05)
return img
async def detect_chart_content(self, img: Image.Image) -> bool:
"""Detect if image contains chart or diagram content"""
# Convert to numpy array
img_array = np.array(img.convert('RGB'))
gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
# Look for geometric shapes (lines, rectangles) typical in charts
edges = cv2.Canny(gray, 50, 150)
lines = cv2.HoughLinesP(edges, 1, np.pi/180, threshold=50, minLineLength=30, maxLineGap=10)
if lines is not None and len(lines) > 10:
# Many lines suggest structured content like charts
return True
# Look for text (charts often have labels)
# This is a simplified check - real implementation would use OCR
text_regions = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
if len(text_regions) > 5:
return True
return False
async def classify_visual_element(self, img: Image.Image) -> str:
"""Classify the type of visual element"""
# This is a placeholder implementation
# Real implementation would use a trained classifier or VLM
if await self.detect_chart_content(img):
return 'chart'
# Check for table-like structure
img_array = np.array(img.convert('RGB'))
gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
# Look for grid patterns
edges = cv2.Canny(gray, 50, 150)
horizontal_lines = cv2.HoughLinesP(edges, 1, np.pi/180, threshold=30, minLineLength=img.width//4, maxLineGap=5)
vertical_lines = cv2.HoughLinesP(edges, 1, 0, threshold=30, minLineLength=img.height//4, maxLineGap=5)
if (horizontal_lines is not None and len(horizontal_lines) > 3 and
vertical_lines is not None and len(vertical_lines) > 3):
return 'table'
# Default to image
return 'image'
async def calculate_relevance_score(self, img: Image.Image, topic: str) -> float:
"""Calculate relevance score of image to presentation topic"""
# Placeholder implementation
# Real implementation would use VLM to analyze image content and compare to topic
base_score = 0.5
# For now, return base score with some random variation
# In real implementation, this would analyze image content
return min(1.0, base_score + np.random.random() * 0.3)
async def generate_image_caption(self, img: Image.Image) -> str:
"""Generate descriptive caption for image using VLM"""
# Placeholder implementation
# Real implementation would use a Vision Language Model
if self.vlm_enabled:
# This would call an actual VLM API
return "AI-generated caption describing the visual content"
else:
return "Visual element extracted from source document"
async def validate_and_filter_visuals(self, visuals: List[VisualElement]) -> List[VisualElement]:
"""Validate and filter visual elements based on quality and relevance"""
validated_visuals = []
for visual in visuals:
# Check quality threshold
if visual.quality_score < self.min_quality_score:
continue
# Check file exists and is readable
if not Path(visual.local_path).exists():
continue
# Check file size (avoid very large files)
if visual.file_size > 10 * 1024 * 1024: # 10MB limit
continue
validated_visuals.append(visual)
return validated_visuals
This visual content extraction system provides comprehensive capabilities for identifying, extracting, and analyzing visual elements from documents. The system ensures that only high-quality, relevant visual content is selected for inclusion in presentations, enhancing the overall quality and effectiveness of the generated slides.
POWERPOINT GENERATION ENGINE
The PowerPoint Generation Engine serves as the culmination of all previous processing steps, transforming the extracted content, structured knowledge, and visual elements into professional presentation slides. This component handles the complex task of organizing information logically, applying consistent design themes, and creating slides that follow UX best practices.
The generation engine implements sophisticated algorithms for content organization, determining optimal slide structures, balancing text and visual content, and ensuring consistent formatting throughout the presentation. It also handles the technical aspects of PowerPoint file creation, including proper XML structure, theme application, and multimedia integration.
The system creates presentations that not only contain relevant information but also follow professional presentation standards with appropriate slide transitions, consistent typography, and effective use of white space. Each slide includes comprehensive notes sections that provide additional context and speaking points.
Here's a comprehensive implementation of the PowerPoint generation system:
The following code example demonstrates how to implement a sophisticated PowerPoint generation engine that creates professional presentations from processed content. This implementation includes slide structure optimization, theme application, content organization, and comprehensive notes generation.
from pptx import Presentation
from pptx.util import Inches, Pt
from pptx.enum.text import PP_ALIGN, MSO_ANCHOR
from pptx.enum.shapes import MSO_SHAPE
from pptx.dml.color import RGBColor
from pptx.enum.dml import MSO_THEME_COLOR
import json
from typing import List, Dict, Tuple, Optional
from dataclasses import dataclass
from pathlib import Path
import re
import asyncio
@dataclass
class SlideContent:
title: str
content_points: List[str]
visual_elements: List[VisualElement]
notes: str
slide_type: str # 'title', 'content', 'visual', 'conclusion'
importance_score: float
@dataclass
class PresentationStructure:
title: str
subtitle: str
slides: List[SlideContent]
theme_name: str
total_estimated_duration: int # minutes
class PresentationGenerator:
def __init__(self, config: Dict):
self.config = config
self.max_slides = config.get('max_slides', 15)
self.max_points_per_slide = config.get('max_points_per_slide', 5)
self.min_words_per_point = config.get('min_words_per_point', 3)
self.max_words_per_point = config.get('max_words_per_point', 20)
# Theme configurations
self.themes = self.load_theme_configurations()
self.default_theme = config.get('default_theme', 'professional_blue')
# UX guidelines
self.ux_guidelines = {
'max_text_per_slide': 50, # words
'min_font_size': 18,
'title_font_size': 32,
'max_bullet_levels': 2,
'visual_text_ratio': 0.6 # 60% visual, 40% text for visual slides
}
async def create_presentation(self, presentation_structure: PresentationStructure,
visual_elements: List[VisualElement],
theme_name: Optional[str],
output_folder: str) -> str:
"""Create PowerPoint presentation from structured content"""
print(f"Creating presentation: {presentation_structure.title}")
# Initialize presentation with theme
prs = self.initialize_presentation_with_theme(theme_name or self.default_theme)
# Create title slide
await self.create_title_slide(prs, presentation_structure)
# Create content slides
for slide_content in presentation_structure.slides:
await self.create_content_slide(prs, slide_content, visual_elements)
# Create conclusion slide
await self.create_conclusion_slide(prs, presentation_structure)
# Apply final formatting and validation
await self.apply_final_formatting(prs)
# Save presentation
output_path = await self.save_presentation(prs, presentation_structure.title, output_folder)
print(f"Presentation created successfully: {output_path}")
return output_path
def initialize_presentation_with_theme(self, theme_name: str) -> Presentation:
"""Initialize presentation with specified theme"""
prs = Presentation()
# Apply theme settings
theme_config = self.themes.get(theme_name, self.themes[self.default_theme])
# Set slide master properties
slide_master = prs.slide_master
# Apply theme colors and fonts
self.apply_theme_to_master(slide_master, theme_config)
return prs
def apply_theme_to_master(self, slide_master, theme_config: Dict):
"""Apply theme configuration to slide master"""
# This is a simplified implementation
# Real implementation would modify the slide master XML
pass
async def create_title_slide(self, prs: Presentation, structure: PresentationStructure):
"""Create title slide"""
# Use title slide layout
title_slide_layout = prs.slide_layouts[0] # Title slide layout
slide = prs.slides.add_slide(title_slide_layout)
# Set title and subtitle
title_shape = slide.shapes.title
subtitle_shape = slide.placeholders[1]
title_shape.text = structure.title
subtitle_shape.text = structure.subtitle
# Apply title slide formatting
self.format_title_slide(title_shape, subtitle_shape)
# Add notes
notes_slide = slide.notes_slide
notes_text_frame = notes_slide.notes_text_frame
notes_text_frame.text = self.generate_title_slide_notes(structure)
def format_title_slide(self, title_shape, subtitle_shape):
"""Apply formatting to title slide elements"""
# Title formatting
title_paragraph = title_shape.text_frame.paragraphs[0]
title_paragraph.font.size = Pt(self.ux_guidelines['title_font_size'])
title_paragraph.font.bold = True
title_paragraph.alignment = PP_ALIGN.CENTER
# Subtitle formatting
subtitle_paragraph = subtitle_shape.text_frame.paragraphs[0]
subtitle_paragraph.font.size = Pt(24)
subtitle_paragraph.alignment = PP_ALIGN.CENTER
async def create_content_slide(self, prs: Presentation, slide_content: SlideContent,
visual_elements: List[VisualElement]):
"""Create content slide based on slide type"""
if slide_content.slide_type == 'visual':
await self.create_visual_slide(prs, slide_content, visual_elements)
else:
await self.create_text_slide(prs, slide_content, visual_elements)
async def create_text_slide(self, prs: Presentation, slide_content: SlideContent,
visual_elements: List[VisualElement]):
"""Create text-based content slide"""
# Choose appropriate layout
if slide_content.visual_elements:
slide_layout = prs.slide_layouts[8] # Content with caption layout
else:
slide_layout = prs.slide_layouts[1] # Title and content layout
slide = prs.slides.add_slide(slide_layout)
# Set title
title_shape = slide.shapes.title
title_shape.text = slide_content.title
self.format_slide_title(title_shape)
# Add content
if len(slide.placeholders) > 1:
content_placeholder = slide.placeholders[1]
self.populate_content_placeholder(content_placeholder, slide_content.content_points)
# Add visual elements if present
if slide_content.visual_elements:
await self.add_visual_elements_to_slide(slide, slide_content.visual_elements[:1]) # Max 1 per text slide
# Add notes
self.add_slide_notes(slide, slide_content.notes)
async def create_visual_slide(self, prs: Presentation, slide_content: SlideContent,
visual_elements: List[VisualElement]):
"""Create visual-focused slide"""
slide_layout = prs.slide_layouts[6] # Blank layout for custom arrangement
slide = prs.slides.add_slide(slide_layout)
# Add title
title_shape = slide.shapes.add_textbox(Inches(0.5), Inches(0.2), Inches(9), Inches(1))
title_frame = title_shape.text_frame
title_frame.text = slide_content.title
self.format_slide_title(title_shape)
# Add visual elements
visual_top = Inches(1.5)
if slide_content.visual_elements:
await self.add_visual_elements_to_slide(slide, slide_content.visual_elements[:2], visual_top)
# Add minimal text content
if slide_content.content_points:
text_points = slide_content.content_points[:3] # Limit text on visual slides
text_shape = slide.shapes.add_textbox(Inches(0.5), Inches(6.5), Inches(9), Inches(1.5))
self.populate_content_placeholder(text_shape, text_points)
# Add notes
self.add_slide_notes(slide, slide_content.notes)
def format_slide_title(self, title_shape):
"""Format slide title"""
title_paragraph = title_shape.text_frame.paragraphs[0]
title_paragraph.font.size = Pt(28)
title_paragraph.font.bold = True
title_paragraph.alignment = PP_ALIGN.LEFT
def populate_content_placeholder(self, placeholder, content_points: List[str]):
"""Populate content placeholder with bullet points"""
text_frame = placeholder.text_frame
text_frame.clear()
for i, point in enumerate(content_points[:self.max_points_per_slide]):
if i == 0:
p = text_frame.paragraphs[0]
else:
p = text_frame.add_paragraph()
# Clean and format point
clean_point = self.clean_bullet_point(point)
p.text = clean_point
p.level = 0
p.font.size = Pt(self.ux_guidelines['min_font_size'])
# Add sub-points if the point is complex
sub_points = self.extract_sub_points(point)
for sub_point in sub_points[:2]: # Max 2 sub-points
sub_p = text_frame.add_paragraph()
sub_p.text = self.clean_bullet_point(sub_point)
sub_p.level = 1
sub_p.font.size = Pt(16)
def clean_bullet_point(self, point: str) -> str:
"""Clean and format bullet point text"""
# Remove excessive whitespace
point = re.sub(r'\s+', ' ', point.strip())
# Ensure proper sentence structure
if not point.endswith(('.', '!', '?', ':')):
point += '.'
# Capitalize first letter
if point:
point = point[0].upper() + point[1:]
# Limit length
words = point.split()
if len(words) > self.max_words_per_point:
point = ' '.join(words[:self.max_words_per_point]) + '...'
return point
def extract_sub_points(self, main_point: str) -> List[str]:
"""Extract sub-points from a complex main point"""
# Look for common sub-point indicators
sub_point_patterns = [
r'including:?\s*(.+)',
r'such as:?\s*(.+)',
r'for example:?\s*(.+)',
r'specifically:?\s*(.+)'
]
sub_points = []
for pattern in sub_point_patterns:
match = re.search(pattern, main_point, re.IGNORECASE)
if match:
sub_text = match.group(1)
# Split on common delimiters
parts = re.split(r'[,;]', sub_text)
sub_points.extend([part.strip() for part in parts if len(part.strip()) > 5])
break
return sub_points[:2] # Limit sub-points
async def add_visual_elements_to_slide(self, slide, visual_elements: List[VisualElement],
top_position: Inches = Inches(1.5)):
"""Add visual elements to slide"""
if not visual_elements:
return
# Calculate positioning for visual elements
slide_width = Inches(10)
slide_height = Inches(7.5)
available_width = slide_width - Inches(1) # Margins
available_height = slide_height - top_position - Inches(0.5)
if len(visual_elements) == 1:
# Single visual element - center it
visual = visual_elements[0]
await self.add_single_visual_element(slide, visual, Inches(0.5), top_position,
available_width, available_height)
else:
# Multiple visual elements - arrange side by side
element_width = available_width / len(visual_elements) - Inches(0.2)
for i, visual in enumerate(visual_elements):
left_position = Inches(0.5) + i * (element_width + Inches(0.2))
await self.add_single_visual_element(slide, visual, left_position, top_position,
element_width, available_height)
async def add_single_visual_element(self, slide, visual: VisualElement,
left: Inches, top: Inches,
max_width: Inches, max_height: Inches):
"""Add single visual element to slide"""
try:
visual_path = Path(visual.local_path)
if not visual_path.exists():
return
# Calculate optimal size maintaining aspect ratio
original_width, original_height = visual.dimensions
aspect_ratio = original_width / original_height
# Determine final size
if aspect_ratio > 1: # Wider than tall
final_width = min(max_width, Inches(original_width / 100))
final_height = final_width / aspect_ratio
else: # Taller than wide
final_height = min(max_height, Inches(original_height / 100))
final_width = final_height * aspect_ratio
# Ensure it fits within constraints
if final_width > max_width:
final_width = max_width
final_height = final_width / aspect_ratio
if final_height > max_height:
final_height = max_height
final_width = final_height * aspect_ratio
# Add image to slide
picture = slide.shapes.add_picture(str(visual_path), left, top, final_width, final_height)
# Add caption if available
if visual.generated_caption or visual.original_caption:
caption_text = visual.generated_caption or visual.original_caption
caption_top = top + final_height + Inches(0.1)
caption_shape = slide.shapes.add_textbox(left, caption_top, final_width, Inches(0.5))
caption_frame = caption_shape.text_frame
caption_frame.text = caption_text[:100] # Limit caption length
# Format caption
caption_paragraph = caption_frame.paragraphs[0]
caption_paragraph.font.size = Pt(12)
caption_paragraph.font.italic = True
caption_paragraph.alignment = PP_ALIGN.CENTER
except Exception as e:
print(f"Error adding visual element {visual.element_id}: {e}")
async def create_conclusion_slide(self, prs: Presentation, structure: PresentationStructure):
"""Create conclusion slide"""
slide_layout = prs.slide_layouts[1] # Title and content layout
slide = prs.slides.add_slide(slide_layout)
# Set title
title_shape = slide.shapes.title
title_shape.text = "Conclusion"
self.format_slide_title(title_shape)
# Generate conclusion content
conclusion_points = self.generate_conclusion_points(structure)
# Add content
content_placeholder = slide.placeholders[1]
self.populate_content_placeholder(content_placeholder, conclusion_points)
# Add notes
conclusion_notes = self.generate_conclusion_notes(structure)
self.add_slide_notes(slide, conclusion_notes)
def generate_conclusion_points(self, structure: PresentationStructure) -> List[str]:
"""Generate conclusion points from presentation structure"""
# Extract key themes from slides
key_themes = []
for slide in structure.slides:
if slide.importance_score > 0.7: # High importance slides
key_themes.append(slide.title)
conclusion_points = [
f"We explored {structure.title} covering key aspects",
f"Main topics included: {', '.join(key_themes[:3])}",
"These insights provide a foundation for further exploration",
"Thank you for your attention"
]
return conclusion_points
def add_slide_notes(self, slide, notes_text: str):
"""Add notes to slide"""
notes_slide = slide.notes_slide
notes_text_frame = notes_slide.notes_text_frame
# Clean and format notes
clean_notes = self.format_notes_text(notes_text)
notes_text_frame.text = clean_notes
def format_notes_text(self, notes_text: str) -> str:
"""Format notes text for speaker notes"""
if not notes_text:
return "No additional notes for this slide."
# Clean up text
notes_text = re.sub(r'\s+', ' ', notes_text.strip())
# Add structure
if not notes_text.startswith("Speaker Notes:"):
notes_text = f"Speaker Notes:\n\n{notes_text}"
# Add speaking tips
notes_text += "\n\nSpeaking Tips:\n- Maintain eye contact with audience\n- Allow time for questions\n- Use gestures to emphasize key points"
return notes_text
def generate_title_slide_notes(self, structure: PresentationStructure) -> str:
"""Generate notes for title slide"""
notes = f"""Speaker Notes for Title Slide:
Welcome the audience and introduce the topic: {structure.title}
This presentation will cover:
- Overview of the subject matter
- Key concepts and insights
- Practical applications and implications
Estimated duration: {structure.total_estimated_duration} minutes
Speaking Tips:
- Start with a compelling hook or question
- Briefly outline what the audience will learn
- Set expectations for interaction and questions"""
return notes
def generate_conclusion_notes(self, structure: PresentationStructure) -> str:
"""Generate notes for conclusion slide"""
notes = f"""Speaker Notes for Conclusion:
Summarize the key points covered in this presentation about {structure.title}
Recap the main themes:
- Reinforce the most important concepts
- Highlight practical applications
- Connect back to the opening objectives
Next steps:
- Encourage questions and discussion
- Provide additional resources if available
- Thank the audience for their attention
Speaking Tips:
- End on a strong, memorable note
- Allow ample time for Q&A
- Be prepared to elaborate on any topic covered"""
return notes
async def apply_final_formatting(self, prs: Presentation):
"""Apply final formatting and validation to presentation"""
# Validate slide count
if len(prs.slides) > self.max_slides:
print(f"Warning: Presentation has {len(prs.slides)} slides, which exceeds recommended maximum of {self.max_slides}")
# Apply consistent formatting across all slides
for slide in prs.slides:
self.validate_slide_content(slide)
def validate_slide_content(self, slide):
"""Validate individual slide content"""
# Check for overly long text
for shape in slide.shapes:
if hasattr(shape, 'text_frame') and shape.text_frame:
text = shape.text_frame.text
word_count = len(text.split())
if word_count > self.ux_guidelines['max_text_per_slide']:
print(f"Warning: Slide has {word_count} words, exceeding recommended maximum")
async def save_presentation(self, prs: Presentation, title: str, output_folder: str) -> str:
"""Save presentation to file"""
# Create output folder if it doesn't exist
output_path = Path(output_folder)
output_path.mkdir(exist_ok=True)
# Generate safe filename
safe_title = re.sub(r'[<>:"/\\|?*]', '_', title)
safe_title = re.sub(r'\s+', '_', safe_title)
filename = f"{safe_title}.pptx"
# Ensure unique filename
full_path = output_path / filename
counter = 1
while full_path.exists():
name_part = safe_title
full_path = output_path / f"{name_part}_{counter}.pptx"
counter += 1
# Save presentation
prs.save(str(full_path))
return str(full_path)
def load_theme_configurations(self) -> Dict:
"""Load theme configurations"""
themes = {
'professional_blue': {
'primary_color': RGBColor(0, 51, 102),
'secondary_color': RGBColor(255, 255, 255),
'accent_color': RGBColor(0, 102, 204),
'font_family': 'Calibri',
'background_style': 'solid'
},
'modern_gray': {
'primary_color': RGBColor(64, 64, 64),
'secondary_color': RGBColor(255, 255, 255),
'accent_color': RGBColor(128, 128, 128),
'font_family': 'Arial',
'background_style': 'gradient'
},
'corporate_green': {
'primary_color': RGBColor(0, 102, 51),
'secondary_color': RGBColor(255, 255, 255),
'accent_color': RGBColor(51, 153, 102),
'font_family': 'Calibri',
'background_style': 'solid'
}
}
return themes
This PowerPoint generation engine creates professional presentations that follow UX best practices, maintain consistent formatting, and include comprehensive speaker notes. The system handles complex content organization while ensuring visual appeal and readability.
COMPLETE WORKING EXAMPLE
Now I'll provide a complete working example that demonstrates how all the components work together to create a functional Agentic AI system for PowerPoint generation. This example includes the main orchestration logic and shows how to use the system from start to finish.
The following complete example demonstrates how to integrate all the components into a working Agentic AI system. This implementation shows the full workflow from user input to final presentation generation, including configuration management, error handling, and user interaction capabilities.
import asyncio
import logging
import json
from pathlib import Path
from datetime import datetime
from typing import Dict, List, Optional
import argparse
# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
class AgenticPowerPointSystem:
"""Complete Agentic AI system for PowerPoint creation"""
def __init__(self, config_file: str = "config.json"):
self.config = self.load_configuration(config_file)
self.initialize_components()
def load_configuration(self, config_file: str) -> Dict:
"""Load system configuration"""
default_config = {
"web_search": {
"max_documents": 15,
"search_engines": ["google", "bing"],
"timeout_seconds": 30
},
"content_processing": {
"min_content_length": 500,
"max_content_length": 50000,
"chunk_size": 512,
"chunk_overlap": 50
},
"rag_system": {
"embedding_model": "all-MiniLM-L6-v2",
"max_chunks_per_query": 10,
"rag_storage_path": "./rag_storage"
},
"graph_rag": {
"min_entity_confidence": 0.7,
"concept_cluster_count": 10,
"graphrag_storage_path": "./graphrag_storage"
},
"visual_processing": {
"min_image_size": [100, 100],
"max_image_size": [1920, 1080],
"min_quality_score": 0.5,
"visual_storage_path": "./visual_content",
"vlm_enabled": False
},
"presentation": {
"max_slides": 12,
"max_points_per_slide": 5,
"default_theme": "professional_blue",
"output_folder": "./presentations"
},
"llm": {
"provider": "openai", # or "local"
"model": "gpt-3.5-turbo",
"api_key": "",
"local_model_path": ""
}
}
config_path = Path(config_file)
if config_path.exists():
try:
with open(config_path, 'r') as f:
user_config = json.load(f)
# Merge configurations
self.merge_configs(default_config, user_config)
except Exception as e:
logger.warning(f"Error loading config file: {e}. Using default configuration.")
else:
logger.info("Config file not found. Creating default configuration.")
with open(config_path, 'w') as f:
json.dump(default_config, f, indent=2)
return default_config
def merge_configs(self, default: Dict, user: Dict):
"""Recursively merge user config into default config"""
for key, value in user.items():
if key in default:
if isinstance(default[key], dict) and isinstance(value, dict):
self.merge_configs(default[key], value)
else:
default[key] = value
def initialize_components(self):
"""Initialize all system components"""
logger.info("Initializing Agentic PowerPoint System components...")
# Initialize components with configuration
self.web_searcher = WebSearchAgent(self.config["web_search"])
self.content_processor = ContentProcessor(self.config["content_processing"])
self.rag_system = RAGSystem(self.config["rag_system"])
self.graph_rag = GraphRAGSystem(self.config["graph_rag"])
self.visual_processor = VisualProcessor(self.config["visual_processing"])
self.presentation_generator = PresentationGenerator(self.config["presentation"])
self.llm_interface = LLMInterface(self.config["llm"])
# Create necessary directories
self.create_directories()
logger.info("All components initialized successfully")
def create_directories(self):
"""Create necessary directories for the system"""
directories = [
self.config["rag_system"]["rag_storage_path"],
self.config["graph_rag"]["graphrag_storage_path"],
self.config["visual_processing"]["visual_storage_path"],
self.config["presentation"]["output_folder"],
"./downloaded_documents"
]
for directory in directories:
Path(directory).mkdir(exist_ok=True)
async def create_presentation_from_topic(self, topic: str, theme: Optional[str] = None,
max_slides: Optional[int] = None) -> str:
"""Main method to create presentation from topic"""
logger.info(f"Starting presentation creation for topic: {topic}")
try:
# Step 1: Search and download documents
logger.info("Step 1: Searching and downloading relevant documents...")
documents = await self.web_searcher.search_and_download(
topic,
max_documents=self.config["web_search"]["max_documents"]
)
if not documents:
raise Exception("No relevant documents found for the topic")
logger.info(f"Downloaded {len(documents)} documents")
# Step 2: Extract and process content
logger.info("Step 2: Extracting and processing content...")
processed_content = await self.content_processor.process_documents(documents)
if not processed_content:
raise Exception("No content could be extracted from downloaded documents")
logger.info(f"Processed {len(processed_content)} content pieces")
# Step 3: Build RAG system
logger.info("Step 3: Building RAG system...")
await self.rag_system.index_content(processed_content)
# Step 4: Create knowledge graph and ontology
logger.info("Step 4: Creating knowledge graph and ontology...")
ontology = await self.graph_rag.create_ontology(processed_content)
# Step 5: Extract visual elements
logger.info("Step 5: Extracting visual elements...")
visual_elements = await self.visual_processor.extract_visuals(documents)
logger.info(f"Extracted {len(visual_elements)} visual elements")
# Step 6: Generate presentation structure
logger.info("Step 6: Generating presentation structure...")
presentation_structure = await self.generate_presentation_structure(
topic, processed_content, ontology, max_slides or self.config["presentation"]["max_slides"]
)
# Step 7: Create PowerPoint file
logger.info("Step 7: Creating PowerPoint presentation...")
output_path = await self.presentation_generator.create_presentation(
presentation_structure,
visual_elements,
theme or self.config["presentation"]["default_theme"],
self.config["presentation"]["output_folder"]
)
logger.info(f"Presentation created successfully: {output_path}")
return output_path
except Exception as e:
logger.error(f"Error creating presentation: {e}")
raise
async def generate_presentation_structure(self, topic: str, processed_content: List[ExtractedContent],
ontology: Dict, max_slides: int) -> PresentationStructure:
"""Generate presentation structure using LLM and processed content"""
# Extract key concepts and entities from ontology
key_concepts = list(ontology.get("concepts", {}).keys())[:10]
key_entities = list(ontology.get("entities", {}).keys())[:15]
# Generate presentation outline using LLM
outline_prompt = self.create_outline_prompt(topic, key_concepts, key_entities)
outline_response = await self.llm_interface.generate_response(outline_prompt)
# Parse outline and create slide structure
slides = await self.create_slides_from_outline(outline_response, processed_content, max_slides)
# Generate title and subtitle
title = await self.generate_presentation_title(topic)
subtitle = await self.generate_presentation_subtitle(topic, key_concepts)
return PresentationStructure(
title=title,
subtitle=subtitle,
slides=slides,
theme_name=self.config["presentation"]["default_theme"],
total_estimated_duration=len(slides) * 2 # 2 minutes per slide estimate
)
def create_outline_prompt(self, topic: str, key_concepts: List[str], key_entities: List[str]) -> str:
"""Create prompt for LLM to generate presentation outline"""
prompt = f"""Create a presentation outline for the topic: "{topic}"
Based on the research, the following key concepts were identified:
{', '.join(key_concepts[:5])}
Key entities mentioned include:
{', '.join(key_entities[:8])}
Please create a logical presentation structure with 8-12 slides that:
1. Introduces the topic clearly
2. Covers the main concepts in a logical order
3. Includes practical examples or applications
4. Concludes with key takeaways
For each slide, provide:
- Slide title
- 3-5 main points to cover
- Slide type (introduction, content, visual, conclusion)
- Importance score (1-10)
Format your response as a structured outline."""
return prompt
async def create_slides_from_outline(self, outline_response: str,
processed_content: List[ExtractedContent],
max_slides: int) -> List[SlideContent]:
"""Create slide content from LLM outline response"""
slides = []
# Parse outline (simplified parsing - real implementation would be more sophisticated)
slide_sections = self.parse_outline_response(outline_response)
for i, section in enumerate(slide_sections[:max_slides]):
# Generate detailed content for each slide
slide_content = await self.generate_slide_content(section, processed_content)
slides.append(slide_content)
return slides
def parse_outline_response(self, response: str) -> List[Dict]:
"""Parse LLM outline response into structured sections"""
# Simplified parsing - real implementation would use more sophisticated NLP
sections = []
lines = response.split('\n')
current_section = {}
for line in lines:
line = line.strip()
if line.startswith('Slide') or line.startswith('#'):
if current_section:
sections.append(current_section)
current_section = {'title': line, 'points': [], 'type': 'content', 'importance': 5}
elif line.startswith('-') or line.startswith('•'):
current_section['points'].append(line[1:].strip())
if current_section:
sections.append(current_section)
return sections
async def generate_slide_content(self, section: Dict,
processed_content: List[ExtractedContent]) -> SlideContent:
"""Generate detailed slide content"""
title = section.get('title', 'Untitled Slide').replace('#', '').strip()
# Get relevant content from RAG system
context = await self.rag_system.generate_context_for_slide(title)
# Generate detailed points using LLM
content_prompt = f"""Create detailed content for a slide titled: "{title}"
Context from research:
{context[:1000]}
Original outline points:
{chr(10).join(section.get('points', []))}
Generate 3-5 clear, concise bullet points that:
- Are informative and accurate
- Use accessible language
- Include specific examples when possible
- Are suitable for a presentation slide
Also generate comprehensive speaker notes for this slide."""
content_response = await self.llm_interface.generate_response(content_prompt)
# Parse response to extract points and notes
points, notes = self.parse_content_response(content_response)
return SlideContent(
title=title,
content_points=points,
visual_elements=[], # Will be populated later based on visual matching
notes=notes,
slide_type=section.get('type', 'content'),
importance_score=section.get('importance', 5) / 10
)
def parse_content_response(self, response: str) -> Tuple[List[str], str]:
"""Parse LLM content response into points and notes"""
lines = response.split('\n')
points = []
notes_started = False
notes_lines = []
for line in lines:
line = line.strip()
if line.lower().startswith('notes:') or line.lower().startswith('speaker notes:'):
notes_started = True
continue
if notes_started:
notes_lines.append(line)
elif line.startswith('-') or line.startswith('•'):
points.append(line[1:].strip())
notes = '\n'.join(notes_lines) if notes_lines else "No additional notes provided."
return points[:5], notes # Limit to 5 points
async def generate_presentation_title(self, topic: str) -> str:
"""Generate presentation title using LLM"""
prompt = f"Generate a clear, professional presentation title for the topic: {topic}. The title should be engaging but not overly creative. Respond with just the title."
response = await self.llm_interface.generate_response(prompt)
return response.strip().replace('"', '')
async def generate_presentation_subtitle(self, topic: str, key_concepts: List[str]) -> str:
"""Generate presentation subtitle"""
if key_concepts:
return f"An overview of {topic} covering {', '.join(key_concepts[:2])}"
else:
return f"A comprehensive overview of {topic}"
async def extend_presentation(self, presentation_path: str, additional_topic: str) -> str:
"""Extend existing presentation with additional content"""
logger.info(f"Extending presentation with topic: {additional_topic}")
# This would load the existing presentation and add new slides
# Implementation would involve loading the PPTX file, analyzing existing content,
# and generating additional slides that complement the existing structure
# For now, return a placeholder
return f"Extended presentation would be saved to: {presentation_path}"
async def modify_presentation(self, presentation_path: str, modifications: Dict) -> str:
"""Modify existing presentation based on user requests"""
logger.info(f"Modifying presentation: {modifications}")
# This would load the existing presentation and apply requested modifications
# such as changing themes, updating content, or reorganizing slides
# For now, return a placeholder
return f"Modified presentation would be saved to: {presentation_path}"
class LLMInterface:
"""Interface for both commercial and local LLMs"""
def __init__(self, config: Dict):
self.config = config
self.provider = config.get("provider", "openai")
if self.provider == "openai":
self.initialize_openai()
elif self.provider == "local":
self.initialize_local_llm()
def initialize_openai(self):
"""Initialize OpenAI client"""
try:
import openai
self.client = openai.OpenAI(api_key=self.config.get("api_key"))
except ImportError:
logger.error("OpenAI library not installed. Install with: pip install openai")
raise
def initialize_local_llm(self):
"""Initialize local LLM"""
# Placeholder for local LLM initialization
# Would use libraries like transformers, llama-cpp-python, etc.
logger.info("Local LLM initialization not implemented in this example")
self.client = None
async def generate_response(self, prompt: str) -> str:
"""Generate response using configured LLM"""
if self.provider == "openai":
return await self.generate_openai_response(prompt)
elif self.provider == "local":
return await self.generate_local_response(prompt)
else:
return "LLM not properly configured"
async def generate_openai_response(self, prompt: str) -> str:
"""Generate response using OpenAI"""
try:
response = self.client.chat.completions.create(
model=self.config.get("model", "gpt-3.5-turbo"),
messages=[
{"role": "system", "content": "You are a helpful assistant that creates professional presentation content."},
{"role": "user", "content": prompt}
],
max_tokens=1000,
temperature=0.7
)
return response.choices[0].message.content
except Exception as e:
logger.error(f"Error generating OpenAI response: {e}")
return "Error generating response"
async def generate_local_response(self, prompt: str) -> str:
"""Generate response using local LLM"""
# Placeholder for local LLM response generation
return "Local LLM response generation not implemented in this example"
def main():
"""Main function to run the Agentic PowerPoint System"""
parser = argparse.ArgumentParser(description="Agentic AI PowerPoint Generator")
parser.add_argument("topic", help="Topic for presentation generation")
parser.add_argument("--theme", help="PowerPoint theme to use", default=None)
parser.add_argument("--max-slides", type=int, help="Maximum number of slides", default=None)
parser.add_argument("--config", help="Configuration file path", default="config.json")
args = parser.parse_args()
async def run_system():
try:
# Initialize system
system = AgenticPowerPointSystem(args.config)
# Create presentation
output_path = await system.create_presentation_from_topic(
args.topic,
theme=args.theme,
max_slides=args.max_slides
)
print(f"Presentation created successfully: {output_path}")
except Exception as e:
logger.error(f"System error: {e}")
print(f"Error: {e}")
# Run the async function
asyncio.run(run_system())
if __name__ == "__main__":
main()
To use this complete system, you would run it from the command line like this:
python agentic_powerpoint_system.py "Artificial Intelligence in Healthcare" --theme professional_blue --max-slides 10
This complete implementation demonstrates how all the components work together to create a sophisticated Agentic AI system that can automatically generate professional PowerPoint presentations from user-specified topics. The system handles the entire workflow from research to final presentation creation while maintaining high quality standards and following UX best practices.
CONCLUSION AND FUTURE CONSIDERATIONS
The Agentic AI system for PowerPoint creation represents a significant advancement in automated content generation, combining multiple AI technologies to create a comprehensive solution that can transform simple user prompts into professional presentations. The system demonstrates how various AI components can work together orchestrally to accomplish complex, multi-step tasks that traditionally require significant human effort and expertise.
The modular architecture ensures that each component can be independently improved and updated without affecting the entire system. This design approach allows for easy integration of new technologies as they become available, such as more advanced Vision Language Models, improved embedding techniques, or more sophisticated knowledge graph algorithms.
Future enhancements could include real-time collaboration features, integration with enterprise knowledge bases, support for multiple languages, and advanced customization options for different industries or presentation styles. The system could also be extended to support other document formats beyond PowerPoint, such as interactive web presentations or PDF reports.
The implementation demonstrates the practical application of cutting-edge AI technologies in solving real-world business problems, showing how autonomous agents can augment human capabilities while maintaining quality and consistency standards. As AI technologies continue to evolve, systems like this will become increasingly sophisticated and capable of handling even more complex content creation tasks.
No comments:
Post a Comment