Vectorless-RAG

A full-stack Retrieval-Augmented Generation (RAG) system that combines PostgreSQL Full-Text Search, BM25 ranking, and a local Large Language Model (Ollama) to provide accurate, context-aware answers from PDF documents.

---

Overview

Vectorless-RAG is designed to demonstrate that effective Retrieval-Augmented Generation can be built using traditional information retrieval techniques instead of relying solely on vector embeddings.

The application allows users to upload PDF documents, retrieve relevant information using hybrid search, and generate grounded responses with source attribution.

---

Features

• Upload and process PDF documents
• Automatic text extraction and chunking
• PostgreSQL Full-Text Search (FTS)
• BM25 keyword ranking
• Hybrid retrieval pipeline
• Local LLM integration using Ollama
• Source-based answer generation
• Modern React frontend
• FastAPI backend
• Docker support
• Evaluation framework for retrieval comparison

---

Tech Stack

Frontend

• React 19
• Vite
• Tailwind CSS
• Framer Motion
• Axios

Backend

• FastAPI
• SQLAlchemy
• PostgreSQL
• pdfplumber
• rank-bm25
• Ollama

Tools

• Docker Compose
• Uvicorn

---

Project Structure

vectorless-rag/
│
├── backend/
│   ├── app/
│   │   ├── api/
│   │   ├── core/
│   │   ├── db/
│   │   ├── models/
│   │   ├── schemas/
│   │   └── services/
│   ├── evaluation/
│   └── requirements.txt
│
├── frontend/
│   ├── src/
│   ├── public/
│   └── package.json
│
├── docker-compose.yml
└── README.md

---

How It Works

1. Upload a PDF document.
2. Extract and clean the document text.
3. Split the text into searchable chunks.
4. Store the chunks in PostgreSQL.
5. Perform Full-Text Search and BM25 retrieval.
6. Rank the retrieved chunks.
7. Build context from the top results.
8. Generate an answer using Ollama.
9. Return the answer along with its source documents.

---

Installation

Clone the repository

git clone https://github.com/Alexrohith/vectorless-rag.git

cd vectorless-rag

Backend

cd backend

python -m venv venv

# Windows
venv\Scripts\activate

pip install -r requirements.txt

uvicorn app.main:app --reload

Frontend

cd frontend

npm install

npm run dev

---

Docker

Run the entire application using Docker Compose.

docker compose up --build

---

API Endpoints

Method	Endpoint	Description
POST	/upload	Upload PDF documents
POST	/query	Ask questions about uploaded documents
POST	/compare	Compare retrieval strategies

---

Future Improvements

• Dense vector retrieval
• Hybrid sparse + dense search
• Cross-encoder reranking
• Multi-document reasoning
• Authentication
• Streaming responses
• Cloud deployment

License

This project is intended for educational and research purposes.