🔍 AI Research Assistant

Ask questions about any PDF — fully local, no paid APIs, no data leaving your machine.

A local RAG (Retrieval-Augmented Generation) pipeline built with BGE embeddings, Qdrant vector search, and TinyLlama via Ollama. Drop in any PDF and start asking questions — everything runs on your machine.

🚀 Quick Start · 🧠 How It Works · 💡 Example Output · 🗺️ Roadmap

---

---

🧠 How It Works

PDF → Text Extraction → Chunking → BGE Embeddings → Qdrant Vector Store
                                                              ↓
                                    Question → Query Embedding → Similarity Search
                                                              ↓
                                             Top Chunks → TinyLlama → Answer

1. PDF Loader — Extracts text page-by-page using PyMuPDF (fitz)
2. Text Chunker — Splits into overlapping token windows using tiktoken
3. Embedder — Generates dense vector embeddings via BAAI/bge-small-en-v1.5
4. Vector Store — Stores and searches embeddings in-memory using Qdrant
5. Answer Generator — Feeds retrieved chunks as context to TinyLlama via Ollama

---

📁 Project Structure

ai-research-assistant/
├── data/
│   └── papers/              # Drop your PDFs here
├── ingestion/
│   ├── pdf_loader.py        # PDF text extraction (PyMuPDF)
│   └── chunking.py          # Token-based overlapping chunker
├── embeddings/
│   └── embedder.py          # BGE embedding model wrapper
├── retrieval/
│   └── vector_store.py      # Qdrant in-memory vector store
├── llm/
│   └── generator.py         # LLM answer generation via Ollama
├── research_assistant.py    # Main pipeline entry point
├── requirements.txt
└── README.md

---

🛠️ Tech Stack

Component	Tool
PDF Parsing	PyMuPDF (fitz)
Tokenization	tiktoken
Embeddings	BAAI/bge-small-en-v1.5 via sentence-transformers
Vector DB	Qdrant (in-memory)
LLM	TinyLlama via Ollama (fully local)
Language	Python 3.10+

---

⚙️ Setup & Installation

Prerequisites

• Python 3.10+
• Ollama installed

1. Clone the repository

git clone https://github.com/overcastbulb/ai-research-assistant.git
cd ai-research-assistant

2. Create a virtual environment

python3 -m venv venv
source venv/bin/activate        # Linux/macOS
# venv\Scripts\activate         # Windows

3. Install dependencies

pip install -r requirements.txt

4. Pull the local LLM

ollama pull tinyllama

5. Add your PDF

mkdir -p data/papers
cp your_document.pdf data/papers/

6. Run the assistant

python3 research_assistant.py

---

💡 Example Output

Loading PDF...
Loaded 15 pages
Chunking text...
Created 36 chunks
Generating embeddings...
Embeddings generated: 36
Storing vectors...
Vector store ready

Question: Explain the transformer architecture

Contexts retrieved: 3
From pages: [3, 2, 2]

Answer:
The Transformer architecture uses stacked self-attention and point-wise,
fully connected layers for both the encoder and decoder...

---

🔒 Privacy First

Everything runs 100% locally:

• No data is sent to any external API
• No OpenAI, no Anthropic, no cloud dependency
• Your documents stay on your machine

---

🗺️ Roadmap

• Support for multiple PDFs in a single session
• CLI arguments (--pdf path/to/file.pdf --question "...")
• Streamlit web UI for non-terminal users
• Persistent vector store (disk-based Qdrant)
• Swap TinyLlama for Groq API (optional, for faster inference)
• Relevance score display alongside retrieved chunks

---

📄 License

This project is licensed under the MIT License.

---

Built for researchers, students, and developers who want to talk to their documents — privately.

Report a Bug · Request a Feature