How Klemens.AI limits hallucinations, or why our platform doesn't make things up

A real problem in business

Anyone who has used ChatGPT or similar AI tools knows this moment: you ask a question, get a fluent, convincing answer, and then it turns out half the information is made up. The AI quotes nonexistent legal sections, cites publications that never existed, or confidently claims something that is simply untrue. We call this phenomenon AI hallucinations . And this isn't a bug that can be fixed with a single update; it's a fundamental characteristic of large language models. These models are trained to generate text that sounds plausible. They have no built-in mechanism to check whether what they say is true.

In a business environment, this isn't a minor inconvenience. When a lawyer relies on AI that cites a nonexistent regulation, or an analyst bases a recommendation on fabricated data, the consequences can be serious. That's why, from the very beginning, at Klemens, we designed the system to minimize hallucinations.

The answer is always based on documents

The fundamental difference between Klemens and a generic chatbot is where the response comes from. A typical chatbot responds based on its "general knowledge," meaning what it has learned through training on billions of websites. The problem is that this knowledge is static, imprecise, and full of contradictions.

Klemens.AI works differently. Before answering a question, he searches exclusively through your documents: regulations, procedures, contracts, reports, knowledge bases. The answer is a synthesis of what he found in specific sources. If information isn't in the documents, Klemens.AI will inform you about it instead of making up an answer.

This approach is called RAG (Retrieval Augmented Generation): the AI model does not answer "off the top of its head," but based on the materials provided to it. It's as if, instead of asking a colleague who "may have heard something about it," we asked an analyst who would first review the relevant documents and only then answer us.

Citing sources: Know where each piece of information comes from

A document-based response alone isn't enough. It's crucial that the user can verify each piece of information.

That's why Klemens includes citations with each answer: references to specific fragments of documents from which the information comes. It's not about a general "source: company regulations," but about a precise connection: this fragment of the answer comes from this specific paragraph of this specific document. This means that users don't have to blindly trust the AI. With a single click, they can check the source and assess for themselves whether the answer is correct. This is a fundamental change in the human-AI relationship: from "I believe it because the AI said so" to "I know it because I see where it comes from."

Hard constraints on AI behavior

In addition to its document-driven architecture, Klemens.AI has built-in behavioral rules that eliminate common sources of hallucinations:

Prohibition of inventing links

AI models love generating URLs that look plausible but lead nowhere. Klemens.AI is strictly prohibited from creating any links, hyperlinks, or website addresses. The only links in responses are those that actually exist in the source documents.

Prohibition of referring to external sources

Klemens.AI doesn't respond with "consult a lawyer" or "check the ministry's website." If the information is in the documents, Klemens provides it. If it's not, he says directly that he hasn't found the answer. He doesn't pretend to know where to look next.

Expert answer, not general

Klemens.AI is designed to respond as an expert who has analyzed the documents. It doesn't repeat generic advice from the internet, but provides specific information from specific sources.

Reanalyze instead of guessing

Imagine this: you're talking to Klemens.AI about work regulations and then you ask about something from a completely different document. A simple chatbot would try to answer based on the previous context of the conversation and would likely make up an answer. Klemens.AI has a built-in conservative reanalysis strategy . If it detects that the question concerns a new topic or that it's unsure of the answer based on the previous conversation, it automatically returns to the documents and searches them anew. The principle is simple: it's better to spend a few seconds re-searching the sources than to risk an incorrect answer.

Minimizing randomness

AI models have a parameter called "temperature." The higher the temperature, the more creative (but also unpredictable) the answers. A high temperature is great for writing poetry, but terrible when you need a precise answer to a legal question.

Klemens applies a low temperature wherever precision matters, especially when classifying content and making decisions. We leave creativity where it's needed (e.g., when creating documents), but when searching for information, we prioritize repeatability and accuracy.

Can hallucinations be eliminated 100%?

Honestly? No. No AI system in the world guarantees 100% accuracy, just as no human is infallible. However, there's a huge difference between "AI that regularly makes things up" and "AI that answers based on documents and allows you to verify any information." Klemens is designed to fall into the latter category. Our approach is multi-layered protection : documents as the single source of truth, citations enabling verification, strict rules of behavior, a conservative reanalysis strategy, and controlled randomness. Each of these layers individually reduces the risk of hallucinations, and together they create a system you can trust in your daily work.