How we make artificial intelligence more reliable

Since writing this article about hallucinatory Large Language Models (LLMs), I have been working with HOWEST's Polytechnic AI Labs. We designed and built a minimal viable product (MVP) to test my hypotheses for reliable LLMs, supporting a Knowledge Management System (KMS). Chatbots aren't the smartest, and the underlying knowledge system is the cause of that. We want to improve those knowledge systems with your help. Send us a text of the protocols or presentations of dishes and wines that you use in your restaurant; we will edit them and then visit you (online is also possible) so you can assess the reliability of the information. Once the system is ready, you may use it for free for a year to, for example, get student workers on board faster and easier. A text and one hour of your time, can you do that for us? Contact us to book your test session via itops@linguafrancaconsulting.eu

A few features of this MVP

Overview of the features on the left side of the KMS

This is the home screen with the various AI functions on the left. More on that in the following pics.

A simple chat, supported by a not-so-simple modelling approach

An example of a chat; nothing special until you see the preparation of this chat.

The combination of knowledge base and user profile already reduces the number of hallucinations.

A general feature that every LLM has: a description of your user profile.

Create and update a knowledge base in an intuitive way

Here you create your own knowledge base and later supplement it with new content if necessary. This goes beyond classic Retrieval Augmented Generation (RAG), because this is also where the magic of our knowledge modelling happens, which we are now testing for its reliability.

We also noticed that many users still write incomplete or poorly formulated prompts, hence this tool.

There are other features, but they are less relevant in the context of this article.

As Dutch speaking readers can see, I gathered information about one specific and widely known knowledge domain: the hospitality industry.

Everyone visits a café or restaurant from time to time, and every hospitality business owner faces problems finding and training staff. A reliable knowledge system can partially solve this problem. The system is only available in Dutch to keep the scope of the training as limited as possible. The intention is to serve multiple knowledge domains in multiple languages ​​later on.

Something about the Artificial Intelligence research

The knowledge elements are provided with multidimensional typologies and metadata, making the system more accurate and ultimately self-learning. We are now starting the next phase and preparing various reward models to support human-in-the-loop learning (HITL). Your help is needed for this, dear hospitality entrepreneur. I would be happy to visit your business and let you test and evaluate the system for at least one hour. In return, you will receive the system free for one year once we go live. Contact us to book your test session via itops@linguafrancaconsulting.eu