SlideSearcher – Intelligent Slide Retrieval Augmented Generation (RAG) System

This repository demonstrates how to set up an AI-powered slide retrieval system that processes PowerPoint presentations, generates detailed AI-based slide descriptions, and enables semantic search through a Flask web interface. It integrates vision-language models, sentence embeddings, and a vector database (ChromaDB) to deliver a complete Retrieval-Augmented Generation (RAG) pipeline for slides.

---

📦 Key Components

• Flask – Web application for interactive slide search
• Micromamba + Python – Lightweight environment management
• ChromaDB – Vector database for embedding and retrieval
• Qwen2.5-VL (Hugging Face) – Vision-language model for slide description
• Sentence Transformers (Hugging Face) – Text embeddings for semantic search
• PDF2Image + LibreOffice – PowerPoint to PNG conversion pipeline
• Gradio + IBMTheme – Custom UI design integration

---

🧰 Prerequisites

Install required system packages (example for macOS):

brew install --cask libreoffice
brew install poppler

For Linux systems (Power or x86):

sudo dnf install git poppler-utils libreoffice

Clone the project repository:

git clone https://github.com/HenrikMader/SlidesSearcher_public.git
cd Slidesearcher_Public

---

🐍 Environment Setup

1. Install Micromamba

cd ~
curl -Ls https://micro.mamba.pm/api/micromamba/linux-ppc64le/latest | tar -xvj bin/micromamba
eval "$(micromamba shell hook --shell bash)"
micromamba --version

2. Create a Python Environment

micromamba create -n rag_env_slides python=3.11
micromamba activate rag_env_slides

3. Install Dependencies

Install project dependencies via pip:

pip install Flask chromadb pydantic_settings sentence_transformers pdf2image accelerate torchvision gradio tqdm transformers

Check installed packages:

pip list

---

🗃️ Build the Vector Database

Navigate to the project directory:

cd ~/Slidesearcher_Public
rm -rf pipeline/db

Convert PowerPoints to image slides:

python pipeline/convert_from_pptx_to_pdf.py

Generate AI-based slide descriptions:

python pipeline/describe_each_pdf.py

Upload the slide embeddings and descriptions to ChromaDB:

python pipeline/upload_descriptions_to_db.py

This process will create or update your vector database in pipeline/db/.

---

⚙️ Run the Web Application

Start the Flask web interface:

python app.py

Access the web UI in your browser:

http://<IP_of_your_machine>:7680

Login credentials:

• Username: power
• Password: power
  (Login credentials can be changed inside templates/login.html file)

---

🧩 Manage the Vector Database

To rebuild or refresh the database, re-run the ingestion scripts:

python pipeline/convert_from_pptx_to_pdf.py
python pipeline/describe_each_pdf.py
python pipeline/upload_descriptions_to_db.py

The database automatically indexes each slide image and its AI-generated description for fast semantic retrieval.

---

🔍 Query the Slide Search System

Once the web app is running, you can:

• Enter natural language queries (e.g., “Show slides about sales trends”)
• SlideSearcher will:
  1. Embed your query using Sentence Transformers
  2. Retrieve top similar embeddings from ChromaDB
  3. Display matching slides and their AI-generated descriptions

Example workflow:

1. Place .pptx files into Files/PPTX_DIR/
2. Run the ingestion scripts
3. Start Flask:

   python app.py

4. Access the web app and start searching.

---

🌐 Web Interface Overview

• index.html – Main search UI with gallery, modal preview, and download options
• login.html – Secure login interface
• main.html – Optional redirect or post-login landing page

Default port: 7680

---

🎨 Custom Theme (IBMTheme)

Defines a unified Gradio-based interface theme:

• Primary color: IBM Blue
• Fonts: IBM Plex Serif & IBM Plex Mono
• Layout: Rounded cards, clear spacing, modern hierarchy

---

📁 Folder Structure

Slidesearcher/
├─ app.py                           # Flask web app entry point
├─ theme.py                         # IBM Gradio theme definitions
├─ pipeline/                        # Processing and database scripts
│  ├─ db/                           # ChromaDB files
│  ├─ config.py                     # Configuration settings
│  ├─ convert_from_pptx_to_pdf.py   # Converts PPTX → PDF → PNG
│  ├─ describe_each_pdf.py          # AI-based slide descriptions
│  └─ upload_descriptions_to_db.py  # Uploads to ChromaDB
├─ Files/
│  ├─ PPTX_DIR/                     # Input PowerPoints
│  └─ IMG_DIR/                      # Output images/descriptions
├─ templates/
│  ├─ index.html
│  ├─ login.html
│  └─ main.html
└─ README.md

---

🗂️ Example Output Directory

Files/
├─ PPTX_DIR/
│  ├─ Sales_Deck.pptx
│  └─ Training_Manual.pptx
└─ IMG_DIR/
   ├─ Sales_Deck/
   │  ├─ slide_1.png
   │  ├─ slide_1.png.desc.txt
   │  └─ slide_2.png.desc.txt
   └─ Training_Manual/
      ├─ slide_1.png
      └─ slide_1.png.desc.txt

---

🤗 Model Sources (Hugging Face)

SlideSearcher retrieves models directly from the Hugging Face Hub:

• Vision Model: Qwen/Qwen2.5-VL-3B-Instruct
• Sentence Embedding Model: all-mpnet-base-v2

You can replace these models in pipeline/config.py using your preferred Hugging Face repositories.

---

🔗 Additional Resources

• ChromaDB Documentation
• Hugging Face Models
• IBM Plex Typeface
• Qwen2.5-VL Model Card