What is RAG
(Retrieval Augmented Generation)?

Since the explosion of ChatGPT, everyone has been talking about artificial intelligence. But if you've ever asked a generic chatbot about your own business, you've probably experienced the disappointment: it invents data, quotes incorrect prices, or confuses your services with a competitor's. That's exactly where RAG comes in.

RAG — Retrieval Augmented Generation — is the technology that allows a chatbot to go beyond "knowing things in general" and instead know your site specifically. Here's a clear explanation, with no unnecessary jargon.

A Simple Definition of RAG

Imagine you're hiring an assistant. You have two options:

• Option A: you hire someone very well-read who has consumed thousands of books, but knows nothing about your specific business.
• Option B: you hire someone who has read all your internal documents, your website, your catalogue and your FAQ — and can give precise answers to your customers.

A classic LLM (GPT-4o, Claude, etc.) is Option A. RAG is Option B. More precisely, it's an architecture that combines both: the language understanding and generation power of a large language model, paired with a knowledge base built from your own content.

RAG breaks down into three words:

• Retrieval = searching for relevant information within your content
• Augmented = the prompt sent to the LLM is enriched with that information
• Generation = the LLM generates a precise, natural-sounding response

How Vector Indexing Works (Plain English)

This is where things get slightly technical — but we'll keep it digestible.

When Chatbot Flow crawls your site, it doesn't "read" your pages the way you do. Instead, it transforms each piece of text into a list of numbers (a vector) that mathematically represents the meaning of the text. This operation is called embedding.

Two texts that discuss the same topic will have similar vectors. "What are your delivery times?" and "How long until I receive my order?" will have closely related mathematical representations, even though the words are different.

These vectors are stored in a specialised database (pgvector, in Chatbot Flow's case). When a visitor asks a question, that question is itself turned into a vector, and the system finds the content from your site whose vector is closest. Those excerpts are then sent to the LLM to build the response.

RAG vs Classic GPT: Why RAG Doesn't Hallucinate

Hallucination is the biggest flaw of LLMs without RAG. When the model doesn't know something, it tends to make up a plausible-sounding answer rather than admitting ignorance. For general use (drafting an email, explaining a concept), that's manageable. For answering your website visitors, it's a disaster.

Consider a vehicle rental website. Without RAG:

Visitor: "Do you offer electric cars?"
LLM-only chatbot: "Yes, we have a great range of electric vehicles, including Tesla Model 3s and Nissan Leafs!" — completely made up.

With RAG:

Visitor: "Do you offer electric cars?"
RAG chatbot: "Based on the information available on our site, our current fleet includes hybrid vehicles. For fully electric options, I'd recommend contacting our team directly." — based on real content.

RAG constrains the LLM to stick to the facts. If the answer isn't in your content, the chatbot says so — and sends you a notification so you can fill in the gap.

Use Cases: E-commerce, Blog, Business Website

E-commerce site

RAG indexes your product listings, return policy, delivery times and FAQs. Your chatbot can instantly answer "How long does delivery take?", "What size should I choose for this jumper?", or "How do I return an item?" — questions that otherwise flood your customer support inbox.

Blog or editorial site

Your content is your expertise. RAG lets the chatbot guide visitors to relevant articles, summarise your stance on key topics, and handle questions you haven't yet covered by honestly saying "I don't have an article on that yet, but you can ask us directly."

Business or services website

This is the most common use case. Pricing, services offered, coverage areas, ordering process, certifications: everything is indexed. The chatbot effectively handles the first 30 minutes of a sales consultation by answering qualification questions.

How Chatbot Flow Uses RAG

Chatbot Flow is built entirely around RAG. Here's exactly how it works in practice:

1. Daily automatic crawl: your site is scanned once every 24 hours. Only modified pages are reprocessed, making indexing fast and cost-efficient.
2. Isolated vector database: each client has their own pgvector instance. Your data is never mixed with another client's.
3. Hybrid search: Chatbot Flow combines vector (semantic) search with keyword search to find the best excerpts. This is more robust than a single approach alone.
4. Additional content: you can enrich the knowledge base with free-text blocks or PDF files (internal pricing guides, documentation, etc.) without publishing them on your site.
5. 100% EU hosting: data never leaves the EU, simplifying GDPR compliance.

Technical FAQ — 5 Common Questions