What is a RAG chatbot? Retrieval-augmented generation explained simply

What is a RAG chatbot?

Short answer

A RAG (retrieval-augmented generation) chatbot answers in two steps: it first retrieves the most relevant passages from your own content, then instructs an AI model to answer only from those passages — citing them. This grounding is what separates a support assistant you can trust from a generic chatbot that improvises.

The two-step loop, concretely

When a visitor asks "do you offer refunds?", a RAG chatbot doesn't ask the model to remember anything. It (1) retrieves: searches an index of your pages and documents for the passages most relevant to refunds, and (2) generates: hands those passages to the model with strict instructions — answer from this context, cite it, and say so if it isn't enough. The model becomes a reader and summarizer of your content, not an author of plausible fiction.

How retrieval actually finds the right passages

Flocci Chat's retrieval is hybrid: vector search (embeddings in PostgreSQL's pgvector, matching by meaning, so "money back" finds your refund policy) combined with full-text search (matching exact words like product names and error codes), then diversity reranking (MMR) so the context isn't five copies of the same paragraph. Content enters that index automatically — an onboarding crawl plus passive learning of each page as real visitors open it, JavaScript-rendered pages included.

Why RAG alone isn't enough: the confidence gate

Retrieval can come back weak — the visitor's question may simply not be answered anywhere on your site. A trustworthy RAG chatbot must notice that. Flocci Chat scores retrieval confidence on every turn and routes low-confidence turns to a clarifying question ("are you asking about credit cards or bank transfers?") or a human handoff with a built-in support ticket — never to a confident guess. Answered turns stream with source citations a visitor can check. More on this in chatbots that cite their sources.

RAG chatbot vs "training a chatbot on your data"

Marketing pages say "train the AI on your website", but almost no product fine-tunes a model on your content — that would be slow, expensive and stale by tomorrow. What good products do is RAG: index your content and reference it at answer time. The practical questions to ask any vendor are: how does content get into the index (can it see JS-rendered pages? does it re-learn changes?), what happens on weak retrieval (clarify/handoff, or bluff?), and are sources shown. Those three questions are the whole trust story.

Frequently asked questions

Is a RAG chatbot the same as a custom GPT?

Similar idea, different depth. A custom GPT references files you upload manually. A production RAG chatbot maintains a live index of your website (crawl + automatic re-learning), scores retrieval confidence per turn, cites sources, and escalates to humans — behaviors uploads-in-a-playground don't give you.

Does RAG completely stop hallucinations?

It removes the main cause — answering from memory — but weak retrieval can still tempt a model to fill gaps. That's why a confidence gate matters: Flocci Chat clarifies or hands off on low-confidence turns instead of answering, and shows citations on the turns it does answer.

What data does a RAG chatbot need?

Whatever your visitors should get answers from: website pages, docs, PDFs, FAQ text. Flocci Chat ingests the website automatically and accepts uploads for the rest.

Is RAG expensive to run?

Retrieval itself is cheap (database search). The metered cost is the model call per answer — which is why Flocci Chat prices per answered message (1 credit) and doesn't charge for retrieval, ingestion or unanswered turns.

Try it on your own website

One script tag. 500 free credits — roughly 500 answered visitor messages. Every feature on, no card. The widget in the corner of this page is the product, running on its own site.

Create your workspace — free