The Info desk: how RAG gives your website a brain

Imagine a new employee starting on day one. You give them all the handbooks, all the FAQs, all the legal documents, all the emails ever sent about your company. You say: "Read this. Tomorrow you start helping customers."

That's what we do with the Info desk. Except Henk reads it in 4 hours instead of 4 weeks. And he forgets nothing.

The problem: AI knows nothing about your business

ChatGPT is trained on the entire public internet. It knows what Pythagoras' theorem is, who Kennedy was, how a Python function works. It knows nothing about your company. Not what your rooms are called, not what your packages cost, not what your return policy is.

If you use it directly on your website, it does two things wrong:

• It makes things up that sound plausible but aren't true ("hallucination")
• It can't cite sources, visitors don't know if the answer is correct

For a law firm or compliance website, that's unusable. One wrong answer and you have a lawsuit.

The solution: RAG

RAG stands for Retrieval-Augmented Generation. In plain language: you first give the AI the relevant context from your knowledge base, then let it answer. That solves both problems:

• No hallucination, the AI has the facts in front of it, it doesn't have to guess
• Source citations, you know which page the answer came from

How does this work exactly? In four steps.

Step 1: Crawl

We read your entire website. Every page, every FAQ, every PDF. For GACSverplichting that was 218 pages, legal text, sector-specific elaborations, FAQ content, examples. We do this with a crawler that respects what you don't want indexed (robots.txt, pages with noindex).

For customers who also want to index content behind logins (like an internal knowledge base), we arrange that via a secure API connection. We don't store anything that doesn't need to be stored.

Step 2: Chunking + Embedding

A 218-page legal text is too big to give to an AI all at once. We cut it into chunks, pieces of about 500 tokens (a few paragraphs). Each chunk gets metadata: which page it came from, which chapter, which section.

Then comes the magical part: embedding. We convert each chunk to a vector, a list of 1,536 numbers that capture the semantic meaning of that text. That sounds abstract. The concrete benefit: chunks about the same topic end up "close to each other" in vector space, even if they use different words.

"How much does it cost?" and "What's the price?" use completely different words. In vector space they sit next to each other. That's why RAG works so much better than old-fashioned keyword search.

Step 3: Retrieve

Visitor asks a question. For example: "Does the GACS obligation also apply to agricultural businesses under 50 employees?"

We embed the question the same way, into a vector. Then we search in our database (Supabase with pgvector) which 5 chunks lie closest. Those are the most relevant pieces from the entire knowledge base.

For this question, Henk finds: a chunk about the general obligation, a chunk about exceptions for agricultural sector, a chunk about the 50-employee threshold, a chunk about deadlines, and a chunk about sanctions.

Step 4: Generate

Now we give these 5 chunks together with the question to Claude (or another LLM). The prompt is roughly:

"Here are 5 relevant pieces from the GACSverplichting knowledge base: [chunks]. Answer the question: '[question]' based on these pieces. Cite the source. If the answer isn't in the pieces, say so honestly."

Claude reads the chunks, thinks, and gives a direct answer with source citation. In our case: "Yes, GACS also applies to agricultural businesses under 50 employees, with an exception for family businesses with fewer than 5 employees (see section 4.2). Deadline for implementation: December 31, 2026 (section 7.1)."

With links to the exact sections. No hallucination. No guessing. Only what's in your knowledge base.

Why this works for content-rich websites

This isn't a theoretical story. For GACSverplichting, the Info desk has:

• 90% of questions answered directly, only 10% require human follow-up
• 3 seconds average response time
• 100% source citations, every statement linked to a legal article
• 0 hallucinations reported over 4 months in production

The reason: GACSverplichting is a content-rich website. Lots of text, lots of detail, lots of cross-references. Exactly where visitors get lost, and where RAG really shines.

Which websites does this work best for?

RAG isn't useful for every site. A 5-page brochure site doesn't need it, a static FAQ is good enough there. RAG shines with:

• Hotels and hospitality, rooms, facilities, policies, conference rooms
• Compliance and legal, regulations with exceptions and deadlines
• E-commerce with 100+ products, technical specs, variants, return policies
• SaaS with extensive documentation, guides, integrations, pricing models
• Education and knowledge institutions, courses, syllabi, regulations

What it costs to build this

Honest about costs. The underlying tech (embeddings, vector database, LLM API) costs us about €60 per month fixed plus €0.02 per conversation variable. Those are pure cloud costs.

The real work is in: crawling the website in a way that respects structure, chunking properly so answers cite correctly, tuning prompts for your specific domain, and updating when your content changes.

We do all that. Live in 5-10 working days. Cancel any month, if it doesn't work, you're not stuck.

The bigger picture

For ten years companies invested in better websites. Prettier designs, faster pages, smarter navigation. All attempts to solve the search problem: how does a visitor find the right information?

RAG removes the search problem. Visitors ask, Henk answers. No menus, no filters, no clicking 4 levels deep. A conversation.

That's why the Info desk changes so much. It's not "adding a chatbot to your site". It's making your site fundamentally different.