For a long time, getting an answer from your own content was a one-way street. A visitor asked a question, the system searched your documents once, and an answer came back. For most questions, that was enough. But the questions people type into a website have changed. They're longer, more specific, and often hold several questions inside one. A single search doesn't always do them justice anymore.

This is the story of how vragen.ai grew from that single, fixed search into something that reasons its way to an answer. Not a different promise, but a better way of keeping it: reliable, traceable answers from your own content.

Reliable answers from your own content

A language model on its own only knows what it learned during training. It knows nothing about your products, your policies, or the document you published last week, and when it doesn't know something, it'll often fill the gap with something that just sounds right. That's the last thing you want on your own website.

This is the problem RAG (Retrieval-Augmented Generation) solves. Before answering, the system first retrieves the most relevant passages from your own content, and only then writes an answer grounded in what it found. The result is an answer you can trust, complete with the sources it came from. It has been the foundation of vragen.ai from the start, and it is why our answers point back to where they came from rather than asking you to take them on faith.

The limit of a single search

In the first generation of vragen.ai, that retrieval ran as a fixed route. A question came in, the system searched once in a pre-set way, and handed the results to the model. Reliable and fast, and for straightforward questions it still works well.

But it had one limitation built into its design: the search happened only once, and always the same way. The model that is good at understanding language never had a say in what to look for. It only ever saw what that first search happened to return.

Think about how you look something up yourself. You search, skim the results, notice your wording was off, and try again. You spot a document worth reading in full instead of a single snippet. You narrow things down: only this year, only this category. A single fixed search can do none of that. If the first attempt misses, the answer suffers, and there is no second chance. The difference that matters is not how good that one search is, but whether the system can search again when the first try falls short.

From a fixed route to a thinking agent

So we stopped fixing the route in advance.

Instead of one search followed by one answer, vragen.ai now works with agents. An agent is given a set of tools and the freedom to decide how to use them. It reasons about the question, chooses whether and what to search, reads what comes back, refines its approach, and continues through several thinking steps until it has enough to answer well. This is what we call agent mode: one question, multiple thinking steps, a better answer.

Importantly, this is not a break with what came before. An agent can begin with a warm start: it first runs a familiar retrieval pass to bring in a handful of likely-relevant documents before it starts thinking. The proven foundation we built for classic search did not disappear; it became one reliable tool the agent reaches for among several. Agents are an evolution of our RAG foundation, not a replacement for it.

That warm start is also why agents are quick. Because this preflight step already pulls in the relevant documents the way our classic pipeline did, an agent answers everyday questions just as fast as the old setup; it doesn't have to go searching from scratch. When a question needs more, you can tell an agent to dig deeper and think longer, giving up a little speed for a better answer where it counts. So the default really is the best of both: as quick as classic search day to day, and ready to go further the moment a question asks for it.

How an agent works through a question

Once it is thinking, an agent moves much the way a careful colleague would. It starts with the documents it has in hand and decides it needs more. It writes its own search phrase, rewrites it when the first attempt comes up short, and breaks a multi-part question into separate searches when that serves the answer better. Where the old route applied one fixed way of processing a query, the agent now makes those calls itself, guided by its instructions. It discovers which filters your content offers, by date, by category, by whatever your documents are tagged with, and applies them to zero in on the right material. When the snippet that best matched the question doesn't actually hold the answer, it opens the full document and reads it properly. Then it brings everything together into a sourced answer.

That kind of back-and-forth is simply not possible on a fixed route. It is the difference between a form letter and a conversation, and it is exactly what makes the hard, multi-part questions answerable.

Everything remains traceable

A more capable system is only worth having if you can still see what it is doing. With vragen.ai, you can. Every step an agent takes is recorded: each search it runs, the filters it applies, the documents it reads, its reasoning along the way, and the sources it ultimately cites. When an answer matters, you can see exactly how the agent arrived at it.

That has been our principle from the very beginning. Not a black box, but a platform you can see into and steer, so you're never left wondering whether the AI quietly made something up in your organisation's name.

Classic search and agents, side by side

Moving everything to agents

Agents are becoming the foundation of vragen.ai, and we're moving existing setups over to them. AI does the heavy lifting of the migration, but we always do it together with you. We look at how your current setup is configured, move it across, and check that it behaves the way you expect before anything changes for your visitors.

The result is a setup that keeps doing exactly what it did before. There's nothing to rebuild, and on day one your visitors get the same answers from the same content. What changes is what becomes possible afterwards: the very same setup can now reason, search again, filter, and read full documents whenever a question calls for it.

The advanced controls you relied on don't disappear either; they simply move to where the agent can act on them. Things like query rewriting and breaking a question into parts used to be fixed steps in the pipeline. Now they live in the agent's instructions, which means the agent applies them with judgement rather than running them the same way every time. You keep the same control over how questions are handled, expressed in a form that can adapt to the question in front of it.

What stays the same is the part that mattered all along: answers drawn from your own content, with the sources to back them up, and full insight into how they were made.