Smart Assistant for Research Summarization

This project is a lightweight AI-powered assistant that reads documents (PDF/TXT), auto-generates summaries, answers questions based on the document, and challenges the user with logic-based questions. It uses NLP models under the hood to provide an interactive, intelligent reading experience.

---

Features

• Upload PDF or TXT documents
• Automatic summarization (≤150 words)
• Ask questions based on the document content
• Challenge Mode: auto-generated logic/comprehension questions
• Answer evaluation with similarity feedback

---

Project Structure

smart_assistant/
├── app.py                # Main Streamlit app logic
├── summarizer.py         # Document summarization
├── qa.py                 # Q&A from document context
├── challenge.py          # Question generation and answer evaluation
├── utils.py              # File reading utilities (PDF/TXT)
├── requirements.txt      # Python dependencies
└── README.md             # You are here!

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/gtbSrihari/Smart-Assistant.git
cd Smart-Assistant

2. Create a Virtual Environment

python -m venv assistant_env
source assistant_env/bin/activate  # or assistant_env\Scripts\activate on Windows

3. Install Requirements

pip install -r requirements.txt

4. Run the Application

streamlit run app.py

---

Architecture & Reasoning Flow

Input Phase

• User uploads a .pdf or .txt file.
• File is processed using pdfminer.six or direct read.

Summarization

• A BART-based transformer model (distilbart-cnn-12-6) generates a concise summary.
• Only the first 4000 characters of the doc are summarized to optimize time.

Question Answering (Ask Anything)

• A DistilBERT QA model (distilbert-base-uncased-distilled-squad) processes the user's question and extracts an answer from the entire document, not just the summary.

Challenge Me (Logic Mode)

• A T5-based model (valhalla/t5-small-qg-hl) generates 3 logic/comprehension questions from the document using keyword-based highlighting.
• The user answers each question.
• Answers are evaluated using cosine similarity (TF-IDF) to check semantic correctness.

---

Notes

• Summarization is capped to 150 words for brevity.
• Question evaluation isn't based on keywords alone but contextual similarity.

---

Future Enhancements

• Long document chunking + multi-turn QA
• Highlighting supporting evidence in document
• Session memory to follow multi-step logic
• Deploy to Hugging Face Spaces

---

Author

Built by Sri Hari Kande – https://github.com/gtbSrihari?tab=repositories

---

License

MIT License – feel free to use, modify, and contribute.