glossary

What Is RAG (Retrieval-Augmented Generation)?

Retrieval-Augmented Generation (RAG) is a technique where an AI first looks something up in a document store, then writes its answer using what it found. It reduces hallucination and lets models cite sources. Perplexity is a pure RAG product; ChatGPT, Claude, and Gemini all use it under the hood for web search and file chats.

5 min read · Updated Apr 17, 2026

The model looks it up before answering

Normally, a language model answers from its internal training — which is outdated, incomplete, and occasionally just wrong. Retrieval-Augmented Generation fixes this by giving the model a “lookup” step:

  1. You ask a question.
  2. A retrieval system searches a document store (your files, the web, a knowledge base) for the most relevant chunks.
  3. Those chunks get inserted into the prompt before the LLM generates an answer.
  4. The model writes a response grounded in what it just read, often with citations.

It’s the difference between a student writing an essay from memory versus writing with the textbook open.

Where you already see RAG in the wild

  • Perplexity — every answer is RAG over a live web search.
  • ChatGPT web browsing and file chats — retrieval before generation.
  • Claude Projects — retrieval across your uploaded documents.
  • Gemini with Workspace — retrieval across your Gmail, Drive, and Docs.
  • NotebookLM — RAG over notebooks you build.

Why this matters for picking a subscription

If your AI work leans toward citable research, document Q&A, or working with private knowledge, RAG-native tools save hours:

  • Citation-first research — Perplexity Pro is built for this.
  • Private document analysis — Claude Projects or ChatGPT Projects on ChatGPT Plus.
  • Workspace-wide knowledge — Gemini AI Pro or Ultra plugged into your Google account.

See our ChatGPT vs Perplexity comparison for how they differ on RAG quality.

RAG’s limits

  • Garbage in, garbage out. If the retriever pulls the wrong chunks, the model writes a confident wrong answer.
  • Still hallucinates. Even with good retrieval, LLMs can over-interpret or misattribute.
  • Chunk boundaries cut context. A sentence split across retrieved pieces can lose its meaning.

Going deeper

Frequently asked questions

What does RAG stand for?

Retrieval-Augmented Generation. The model retrieves relevant information from an external source, then generates its response using that information.

Is Perplexity using RAG?

Yes — Perplexity is essentially a RAG product with a polished UI. It retrieves web pages in real time, then an LLM writes a cited answer based on them.

Does ChatGPT use RAG?

Yes, when you use web browsing, file upload chats, or custom GPTs with knowledge files. Under the hood, the relevant chunks are retrieved and inserted into the prompt before generation.

How is RAG different from fine-tuning?

Fine-tuning bakes knowledge into the model itself; RAG looks it up at query time. RAG is cheaper, updates instantly when documents change, and lets the model cite sources — but it depends on retrieval quality.

Why does RAG reduce hallucinations?

Because the model is working from actual source text, not its memory. It's much harder to invent details when the relevant passage is literally in the prompt.

Can I build RAG into my ChatGPT or Claude subscription?

Sort of. Custom GPTs, Projects, and Claude Projects all act as lightweight RAG setups over files you upload. For serious workloads, developers use vector databases and the API.