Why is Minerva not just another LLM?

AILLMTechnology

Minerva Team

The Minerva product and content team.

The moment when LLMs gained widespread popularity came at the turn of 2022 and 2023, when generative artificial intelligence entered mainstream use. These models quickly drew attention thanks to their ability to provide detailed information on a wide range of topics. However, after the initial excitement, their fundamental weaknesses also became apparent — namely, the scale of incorrect answers.

According to the Tom’s Guide report, ChatGPT-5 generates incorrect responses in roughly 25% of cases. This means that 1 in 4 answers can be wrong.

Are LLMs not reliable sources of information?

Even though language models such as ChatGPT or Gemini aim to deliver precise answers, they often make mistakes because their responses are based on patterns rather than on verified facts and data. In IT terminology, this phenomenon is referred to as “artificial intelligence hallucination.” These results often stem from internal model flaws, a lack of real-world understanding, or limitations in training data. In other words, the AI system "hallucinates" information that was never explicitly provided during training, which leads to unreliable or misleading output.

The term hallucination in relation to artificial intelligence refers to generating outputs that may sound convincing but are either factually incorrect or unrelated to the given context.

Why is Minerva not just another LLM?

While Minerva leverages engines such as ChatGPT and Gemini, equating it with a generic language model (LLM) is a mistake. The fundamental difference lies in data verifiability and accountability in the decision-making process.

With a standard AI chat, the user receives an answer that may suffer from hallucination — the model can “make up” information that is not present in the text. Minerva eliminates this risk by providing not only an answer to the key question, but also a specific quotation along with a link that leads directly to the relevant place in the tender documentation, enabling immediate fact-checking.

As a result, the user can be certain that the information is not randomly generated but originates from the analyzed file.

Another distinguishing factor is Minerva’s deep understanding of industry-specific context — something generic LLMs often struggle with. Tender language is highly specific and non-standardized: the same requirements, such as bid bonds or performance guarantees, may be written in countless different ways. A typical AI model may misinterpret these nuances or overlook key information hidden under unusual wording.

Minerva can connect facts contextually: for example, it understands that a tender titled “road modernization” may in reality concern building a safe pedestrian crossing — which is crucial for companies with narrow specialization.

A significant advantage of Minerva over “manual” use of tools like ChatGPT is the specific quotation along with a link. Instead of forcing the user to type dozens of keywords and manually search hundreds of pages for mentions of a specific product (e.g., a scalpel or a surgical bed), the system analyzes the client’s portfolio and automatically delivers tenders in which those products appear. It works as a form of “reverse search”: the user is no longer looking for a needle in a haystack; the system shows only those procedures that contain solutions from their offering. This allows for initial filtering and saves resources even before a deep dive into documentation.

Security within the Minerva ecosystem

Implementing Minerva is also a step toward process standardization and data security — something that individual use of public AI models by employees cannot guarantee. When using ChatGPT on their own, each employee may apply different, often incorrect prompts and may also expose the company to data leakage by uploading internal files to a public chat.

Minerva creates a closed, secure ecosystem in which the analysis process is repeatable and standardized across the entire team. This prevents the chaos of each salesperson searching and analyzing tenders in a different way, and ensures that decisions about whether to participate in a procedure are made based on reliable, filtered data.