Why ezrag?

Most RAG stacks take hours to wire up: embedding models, vector stores, chunkers, prompts, and LLM clients. ezrag hides all of that behind one class, so you go from an empty folder to a grounded, citation-aware answer in exactly three lines of code.

What you get out of the box:

• Chunking — sentence-aware splitting with a sliding overlap, zero configuration
• Embeddings — a lightweight sentence-transformers model, loaded lazily
• Vector DB — a self-contained in-memory store with fast NumPy cosine search
• Context bundling — evidence is ranked, tagged [1] [2] [3] ... and injected into the prompt
• LLM execution — unified calls through the Hugging Face Serverless API or LiteLLM
• Weak-evidence guard — answers truthfully when the documents do not cover the question

Installation

pip install ezrag

Quick Start

That is the entire experience — import, load, ask:

from ezrag import EZRAG                     # 1. import

rag = EZRAG().load(SOURCE)                  # 2. initialise + load your source
print(rag.ask("Your question here"))        # 3. ask — grounded, cited answer

No environment variables to set up, no vector database to spin up, no prompting prompts to craft. One object call chain produces a grounded, cited answer.

Configuration

Providers

• Hugging Face Serverless (default) — no custom API wiring. Set the HF_TOKEN environment variable only when the model is gated.
• LiteLLM — point llm at any OpenAI-compatible endpoint, for example openai/gpt-4o or azure/gpt-4o.

How It Works

  source ──▶ chunk ──▶ embed ──▶ in-memory index
                                       │
  question ──▶ embed ──▶ top-k search ──┘
                                       │
                      context bundle ──▶ LLM ──▶ cited answer

Hugging Face Spaces

ezrag runs out of the box in a Hugging Face Space — pip pulls in every runtime dependency automatically. Store a HF_TOKEN Space secret if your chosen model is gated.

License

MIT © Oguru Vinay Reddy