PDF AI Assistant (CLI-based RAG System)

A command-line AI assistant that lets you embed and chat with PDF documents using OpenAI embeddings, pgvector, and GPT-4.

---

✨ Features

• Parse and clean PDFs using PyMuPDF
• Generate OpenAI text-embedding-3-small vector embeddings
• Store embeddings in PostgreSQL with pgvector
• Retrieve semantically similar pages using cosine similarity
• Chat with your PDF using GPT-4 (or GPT-4o), with streamed responses
• Interactive CLI with multi-document support

---

📦 Requirements

• Python 3.10+
• PostgreSQL with pgvector extension enabled
• OpenAI API key

Python Dependencies

pip install -r requirements.txt

---

⚙️ Setup

Database Setup:

1. Create a PostgreSQL database

2. Install the pgvector extension:

   CREATE EXTENSION vector;

3. Configure your database connection in app/core/postgres.py

OR

Just run

docker compose up

at the root folder.

Environments:

POSTGRES_DB -> The database name to be created or used

POSTGRES_URL -> The database url, such as localhost:5432

POSTGRES_USER -> Database user name

POSTGRES_PASSWORD -> Database password

OPENAI_KEY -> The API Key to be generated from OpenAI Dashboard

---

🚀 Usage

Embed a PDF:

python main.py embed path/to/file.pdf

Chat with your PDFs:

python main.py chat

You’ll be shown a list of available embedded PDFs to choose from. Once selected, type your questions and get AI-powered answers based on document context.

---

💡 Architecture Overview

1. PDF Parsing → Clean each page with PyMuPDF
2. Embedding → Use OpenAI to embed each page’s content
3. Storage → Insert into PostgreSQL with pgvector
4. Retrieval → Find top-N most relevant pages via cosine similarity
5. Prompt Assembly → Inject pages as context + user prompt
6. Response → Stream answers from OpenAI GPT-4o

---

📄 License

MIT — use freely, attribute if you build on it.

---

🙌 Author

Umut Cevdet Koçak

Built as a weekend project to explore vector search, embeddings, and conversational RAG systems.