# What is RAG? Source-grounded AI explained

By the GO AI Team · August 21, 2026

> **TL;DR:** **RAG** — Retrieval-Augmented Generation — means the AI looks something up before answering. It searches *your* documents, puts the relevant passages in front of the model, and asks it to answer from those, citing sources. It fixes training cutoffs, private-document blindness, and much confident invention. Not all of it.

## The problem

A language model is a very good compression of text it read once. Hence fluency, breadth, and three weaknesses: it stops knowing things after its cutoff, it never saw anything private, and when it doesn't know it produces a plausible answer anyway — because producing plausible text is the job.

You can't retrain per document. So you change *where the answer comes from*.

## The four steps

1. **Index** — documents split into passages, each converted to a numeric representation of meaning. Once, up front.
2. **Retrieve** — your question converted the same way; closest passages found. Not keyword matching: "what did we agree about late delivery" can find "shipment delays."
3. **Augment** — those passages go into the model's context with your question and an instruction to answer from them.
4. **Generate and cite** — the model answers, and each claim can point back at its source.

Step 2 is the clever part. Step 4 is the trustworthy part.

## Why "just paste it in" isn't the same

Attaching a document works for one paper. It stops working at ninety of them, when you'll ask again next month, or when the relevant paragraph is on page 340 and the other 339 compete for attention. RAG indexes once, retrieves per question.

## Where RAG still fails

- **Retrieval misses** — the model answers confidently from what it *was* given. Synthesis across scattered passages is the classic hard case.
- **The answer drifts past the evidence** — citations catch this; an uncited claim inside a cited answer deserves suspicion.
- **The source is wrong** — grounding means faithful to your documents, not to reality.
- **Chunking artefacts** — tables split across passages, clauses depending on a definition forty pages earlier.

## When it's worth the setup

| Situation | Use |
|---|---|
| One document, one question | Paste into a chat |
| One long document, queried repeatedly | Either, leaning RAG |
| Dozens of documents on a topic | RAG |
| Answers someone else relies on | RAG — for the citations |
| Mixed media: PDF, video, audio, web | RAG, if it handles all of them |
| General knowledge | Neither — just ask a model |

## As a product

[Neurobase](https://goaichat.app/neurobase) is our version: give a *Neuron* sources — PDF, text, video, YouTube, web pages, images, audio — and it becomes a focused assistant answering with citations. Test in the Lab, chain Neurons into workflows with Summarize, Compare, Extract, Translate.

Concrete version: turn a folder of PDFs into an assistant.

---

We build [Neurobase](https://goaichat.app/neurobase) — source-grounded assistants with citations, for iPhone, iPad, and Mac.

More: [GO AI Blog](https://goaichat.app/blog) · [support@goaichats.com](mailto:support@goaichats.com)