Conclusions

Information retrieval technology has advanced greatly, and current systems continue to improve rapidly. However, there are still technical barriers and constraints to implement fully functional and deployable systems. Training custom neural network (e.g. LLMs) from scratch give the best possible results, but they require large data, specialised hardware, technical human expertise, and they are economically costly. Thus, the most common use of this technology is based on the subscription model with companies such as OpenAI (e.g. ChatGPT).

Given the constraint mentioned above, one common alternative is to use LLMs that have been pre-trained and optimised to carry out tasks within a specific context. Then, these models can be used to train our models as a downstream task.

There are open issues such as privacy, copyrights and IP, bias, and other risks that need to be assessed when implementing AI systems, particularly when new legal frameworks such as the EU AI Act (Regulation (EU) 2024/1689) will affect the deployment of AI systems.