Univox

Univox is an intelligent AI-powered study assistant. It leverages advanced natural language processing, computer vision, and graph-based reasoning to help students learn more effectively through document analysis, content extraction, and personalized support.

The system is built on top of LangGraph, enabling a modular, agent-driven architecture where different AI components collaborate seamlessly. With Univox, students can interact naturally — through text or voice — and access a wide range of academic resources in an intuitive way.

Retrieval-Augmented Generation

A core feature of Univox is its RAG pipeline, which enhances responses with relevant documents retrieved from the student’s course materials:

• Document retrieval: Every query is matched against a FAISS-based vectorstore built from parsed syllabi, books, slides, notes, and multimedia.
• Context enrichment: Retrieved documents are injected into the LLM prompt, ensuring answers are grounded in the actual study material.
• Downloadable resources: In addition to enriched answers, students are given direct links to download useful files (e.g., lecture slides, exam papers, reference books).

This means Univox is not just a conversational agent — it acts as a personal knowledge navigator, combining semantic search with LLM reasoning to maximize learning.

Project Structure

study-buddy/
│
├── data/                   # Raw (only locally), metadata, and processed datasets
│   └── README.md
│
├── faiss_index/            # Vectorstore index
│
├── images/
│
├── study_buddy/            # Core source code
│   └── README.md
│
├── tests/
│   ├── performance/
│   │    └── README.md
│   └── tool_tests/         # Unit and load tests
│
├── streamlit_frontend.py   # Streamlit web app entrypoint
├── config.yaml             # Configurable models, embeddings, vectorstore
├── langgraph.json
├── pyproject.toml          # Project metadata and dependencies
├── requirements.txt
├── setup.cfg
├── .gitignore
├── LICENSE
└── README.md

What You Can Do with Univox

Univox acts as your personal academic companion, combining course-specific knowledge with powerful AI tools to assist you throughout your learning journey.

Course-Specific Assistance

Need quick access to important information? Univox can:

• Provide contact details for professors and teaching assistants
• Show exam dates, office hours, and course schedules
• Extract information directly from syllabi, announcements, and handouts

For example:

“When are the midterm exams for this course?”

“How can I contact Professor Lops?”

Exam Preparation and Practice

Univox helps you study smarter, not harder:

• Retrieve past exam questions and exercises from your uploaded materials
• Generate custom practice problems tailored to your curriculum
• Suggest relevant topics and resources for your upcoming tests

For example:

“Give me practice problems for cosine similarity” “Show me past exams of MRI course”

Topic Clarification and Learning Support

Stuck on a concept? Univox explains it clearly and links you back to the exact source:

• Summarizes difficult topics from lecture notes
• Highlights key formulas and definitions
• Points you to the relevant pages and documents

For example:

“Explain mean reciprocal rank in simple terms”

“What’s the difference between item based and user based recommender systems?”

Advanced Research and Analysis

Univox isn’t just a Q&A bot — it’s also a research companion:

Document Processing

• Extracts text from scanned PDFs and images using Tesseract OCR
• Summarizes long research papers and textbooks
• Handles multiple document formats effortlessly

Data Analysis & Visualization

• Analyzes datasets and generates insights
• Creates visualizations from CSV files using natural language
• Runs custom Python code for statistical tasks

Research Enhancement

• Searches academic sources like ArXiv, Google Scholar, and Wikidata
• Finds books, papers, and supplementary materials
• Accesses up-to-date information via web search

Voice Interaction and Accessibility

Univox also offers hands-free interaction:

• Ask questions using voice commands
• Receive audio-based responses
• Transcribe recorded lectures into searchable text (supports Italian)
• Convert summaries and explanations into natural-sounding audio

This makes Univox highly useful for students with visual impairments, reading difficulties, or those who prefer multimodal learning.

Interactive Learning Experience

• Ask natural language or voice questions about your uploaded materials
• Get direct citations with every answer
• Manage and browse your content via an intuitive Streamlit-based web app
• Integrate multimedia: process text, images, audio, and datasets seamlessly

Transparency and Source Attribution

Univox is designed to be trustworthy:

• Every answer includes clear references to the source documents
• Displays page numbers and relevant sections where possible
• Ensures traceability so you can always verify information

---

Technologies Used

• Python 3.12
• LangGraph for orchestrating multi-agent workflows
• Tesseract OCR for document text extraction
• Configurable Embedding Models — default: BAAI/bge-m3
• Together.ai for LLM integration
• FAISS for fast vector indexing
• Transformers for model inference
• TensorFlow / PyTorch with CUDA acceleration
• Streamlit for the interactive web interface

---

Prerequisites

• Windows 11 64-bit (tested environment)
• Python 3.12
• NVIDIA GPU with CUDA support (RTX 5080 recommended)
• 8GB+ RAM
• ~5GB disk space for models and dependencies
• A Together.ai API key

---

Installation

1. Install Tesseract OCR

1. Download the installer: Tesseract OCR – UB Mannheim

2. Install it in the default directory.

3. Add these paths to System Environment Variables:

   C:\Program Files\Tesseract-OCR
   C:\Program Files\Tesseract-OCR\tesseract.exe
   

4. Verify installation:

   tesseract --version

---

2. Set Up the Python Environment

git clone https://github.com/npinto97/univox.git
cd univox

python -m venv venv
venv\Scripts\activate

pip install --upgrade pip
pip install -r requirements.txt

---

3. GPU Configuration (Optional but Recommended)

Check CUDA availability:

nvidia-smi

Then install the correct CUDA and cuDNN versions for your GPU. For RTX 5080, CUDA 12.x is recommended. See NVIDIA CUDA downloads for more details.

---

4. Initialize the System

1. Parse your course metadata:

   python parse_course_metadata.py

2. Build the FAISS index:

   python update_faiss_index.py

---

Configuration

The system uses config.yaml to customize LLM, embeddings, and vector stores:

llm:
  model: "meta-llama/Llama-3.3-70B-Instruct-Turbo"

embeddings:
  model: "BAAI/bge-m3"

vector_store:
  type: "faiss"

You can replace the model names with any alternative available on Together.ai or Hugging Face.

---

Usage

Start the Streamlit app:

streamlit run streamlit_frontend.py

By default, it launches at: http://localhost:8501

Workflow:

1. Upload your study materials
2. Ask questions via chat or voice
3. View responses with source citations

---

Future Enhancements

• More lightweight embedding models for CPU-only setups
• Enhanced semantic search and document summarization
• Richer multimodal learning experience