Artificial intelligence is becoming increasingly integrated into the digital processes of companies, public institutions, and organizations. However, the real challenge is not simply building ever more powerful models, but designing systems that are reliable, secure, and genuinely useful for people.
This is precisely the focus of the bachelor’s thesis by Giovanni Battista Perini, a student in Multimedia Communication and Information Technologies at the University of Udine, who chose to explore the design of a conversational chatbot based on advanced language models.
The research and development were carried out at emmecubo, a web agency based in Treviso where Giovanni Battista collaborates, providing a perfect example of the synergy between academic education and business innovation.
As part of his thesis, titled The Retrieval Augmented Generation (RAG) Approach: Theoretical Analysis and Implementation Using the Cheshire Cat Framework, Regolo was used as the infrastructure for accessing language models, becoming a technological component of the Oracolo project, a digital assistant designed to provide information to citizens and visitors through simple and natural interaction.
This is a concrete example of how university research can become a testing ground for experimenting with new applications of artificial intelligence.
Oracolo: a digital assistant designed for the local community
The project was born from the idea of using conversational artificial intelligence to improve access to public information.
Oracolo was designed as a chatbot that can be integrated into digital information systems accessible through kiosks installed in urban areas and via a mobile promotional vehicle.
The goal is to enable citizens and visitors to obtain information through natural conversation, without having to navigate through numerous pieces of content or complex interfaces.
To achieve this, the project required careful design of the entire technological architecture: from the selection of the language model to information management, as well as the systems needed to ensure security and reliability.
Why use a RAG approach
One of the central aspects analyzed in the thesis concerns overcoming one of the main limitations of generative language models: their tendency to produce incorrect or outdated responses.
For this reason, Oracolo was developed using an architecture based on Retrieval-Augmented Generation (RAG).
This approach allows the language model to rely not only on its internal knowledge, but also to retrieve information from a dedicated database and use it as context to generate more accurate responses.
In practice, the model is not used as a simple source of knowledge, but as a tool capable of interpreting, organizing, and communicating information drawn from carefully selected sources.
This is a fundamental choice for applications where accuracy and reliability are essential requirements.
Regolo as the infrastructure for experimentation
To develop the chatbot, the thesis analyzed several language models available through Regolo, evaluating characteristics such as comprehension capabilities, response quality, processing speed, and adaptability to the context.
The choice of infrastructure was not merely a technical matter, but also a design decision.
In a scenario intended for a public service, aspects such as data management, regulatory compliance, and control over the technological infrastructure become essential.
Model selection: finding the right balance between performance and reliability
During development, several language models were compared by analyzing their behavior within the same chatbot configuration.
The experimentation showed that model selection does not depend solely on the number of parameters or computational power, but on achieving the right balance between reasoning capabilities, response speed, and the ability to follow the instructions defined by the system.
Among the models tested, the final choice was Llama-3.3-70B-Instruct, which was evaluated as the model best suited to Oracolo’s requirements thanks to its ability to understand natural language, maintain consistency throughout conversations, and work effectively with the context provided by the RAG system.
The semantic search component was also analyzed, leading to the selection of Qwen3-Embedding-8B to improve the retrieval of the most relevant information.
Regolo provided access to multiple language models through a dedicated inference service, offering the opportunity to experiment with open-weight solutions within an environment designed for professional applications.
Building a chatbot also means designing for security
One of the most interesting aspects of the thesis concerns the topic of security.
A digital assistant intended for use in public spaces must be designed with complex scenarios in mind: different users, unpredictable requests, and potential attempts to manipulate the model’s behavior.
For this reason, Oracolo was developed using a multi-layered approach.
Dedicated logic was introduced to manage the chatbot’s behavior, along with systems to reduce the risk of incorrect responses and protection mechanisms against phenomena such as prompt injection and jailbreak attacks – that is, techniques through which a user may attempt to alter the model’s internal instructions.
The experimentation demonstrated a fundamental principle in the development of modern AI: an advanced language model must always be accompanied by a system designed to ensure its controlled and responsible use.
Artificial intelligence as a tool in the service of people
The work developed in the thesis represents a concrete example of how artificial intelligence can be applied to real-world problems.
Oracolo was not created as a simple chatbot, but as a meeting point between technology, the local community, and digital services.
The research demonstrated that building an effective AI assistant requires interdisciplinary expertise: software development, data management, knowledge of language models, and careful attention to regulatory and ethical aspects.
From university research to digital innovation
We are proud that Regolo was chosen as part of a university project dedicated to the study and experimentation of applied artificial intelligence.
Collaboration between universities, research, and technology is essential to building a future in which AI is increasingly accessible, secure, and useful.
Projects such as Oracolo demonstrate that artificial intelligence is not merely a technology to observe, but a tool to be designed thoughtfully in order to create new, tangible opportunities for citizens, businesses, and institutions.
FAQ
What is Oracolo?
Oracolo is a digital assistant that uses conversational AI to provide public information to citizens and visitors through natural language. It’s accessible via kiosks in urban areas and a mobile promotional vehicle.
What is RAG and why was it used in this project?
Retrieval-Augmented Generation (RAG) is an architecture where the language model retrieves relevant documents from a dedicated database and uses them as context to generate grounded, accurate responses. It was chosen to overcome the hallucination problem inherent in generative models.
Which language model powers Oracolo?
The system uses Llama-3.3-70B-Instruct as the generative model and Qwen3-Embedding-8B for semantic search and retrieval, both accessed through Regolo’s inference infrastructure.
What security measures were implemented?
Oracolo uses a multi-layered security approach including behavior-governing logic, hallucination mitigation systems, and protection against prompt injection and jailbreak attacks.
Start your free 30-day trial at regolo.ai and deploy LLMs with complete privacy by design.
👉 Talk with our Engineers or Start your 30 days free →
- Discord – Share your thoughts
- GitHub Repo – Code of blog articles ready to start
- Follow Us on X @regolo_ai
- Open discussion on our Subreddit Community
Built with ❤️ by the Regolo team. Questions? regolo.ai/contact or chat with us on Discord