Stateless RAG

RAG without vector databases — uses tree-indexed PDF navigation + Gemini 2.0 Flash.

---

What It Does

Most RAG systems chunk documents into vector embeddings and run similarity search. This project skips all that.

Instead:

1. PDF is uploaded to PageIndex, which builds a tree index (like a smart Table of Contents)
2. On each query, Gemini navigates the tree to find the right sections
3. Gemini then generates a cited answer from those sections

No embeddings. No vector store. Just two API calls per query.

---

How It Works

PDF → PageIndex (tree index) → User Query
                                    ↓
                          Gemini: "Which nodes answer this?"
                                    ↓
                          Fetch those sections
                                    ↓
                          Gemini: Generate cited answer

---

Tech Stack

What	Tool
Document Indexing	PageIndex API
Language Model	Gemini 2.0 Flash
Interface	Jupyter Notebook
Language	Python 3.11+

---

Setup

1. Clone

git clone https://github.com/Shivas-A54/Stateless_RAG.git
cd Stateless_RAG

2. Install

pip install pageindex python-dotenv google-generativeai ipykernel

3. Add API keys

PAGEINDEX_API_KEY="your_key"   # → dash.pageindex.ai/api-keys
GEMINI_API_KEY="your_key"      # → aistudio.google.com/app/apikey

4. Run

jupyter notebook PageIndex.ipynb

---

Project Structure

Stateless_RAG/
├── PageIndex.ipynb        # Full pipeline
├── RAG.pdf                # Sample PDF for testing
├── workflow_diagram.png   # Architecture diagram
├── .env.example           # API key template
└── pyproject.toml

---

Why "Stateless"?

No vector store is created or maintained locally. The document tree lives in PageIndex (external), and each query runs as a clean, self-contained pipeline. Nothing is written to disk between runs.

---

Limitations

• PDF only (other formats untested)
• Requires a PageIndex API account
• Two Gemini calls per query (adds slight latency)
• Very large PDFs may increase token usage

---

Acknowledgments

• PageIndex — tree indexing API
• Google Gemini — language model

---

Built by Shivas-A54