Interactive Demo

RAG Pipeline

Retrieval-Augmented Generation grounds LLMs in your data. Watch each stage — from raw documents to a cited answer.

Strategy

Recursive Sentence Fixed

Load

Source

Chunk

Split

Embed

Vectorize

Store

Index

Retrieve

Generate

Answer

Press Run Pipeline to watch RAG in action.

What is RAG?

RAG lets an LLM answer questions using your own documents instead of relying on training data. It retrieves relevant passages first, then generates a grounded response.

This eliminates hallucination for factual queries and keeps answers up-to-date without retraining.

The 6 Stages

1 Load — ingest raw documents (PDF, HTML, etc.)
2 Chunk — split into overlapping segments
3 Embed — convert chunks to vectors
4 Store — index in a vector database
5 Retrieve — find nearest chunks to query
6 Generate — LLM answers with context

Key Concepts

Cosine similarity measures how close two vectors are in meaning — not word overlap.

Top-k retrieval returns the k most relevant chunks. Usually k=3-5.

Chunk overlap prevents losing context at split boundaries.

Stack

LangChain OpenAI ChromaDB HNSW t-SNE

results = vectorstore.similarity_search( query, k=3 ) prompt = f"Answer using ONLY:\n{context}"