Retrieval-Augmented Generation (RAG) is a technique that combines generative language models (LLMs) with document retrieval systems.
During generation, the model queries a knowledge base external to the LLM to retrieve relevant information in real time.
This information is then integrated into the response, ensuring accuracy and up-to-date content.
RAG is useful in contexts where it is necessary to respond with up-to-date or specialized data not present in the LLM model that manages communication with the user.
In particular, GAIA is equipped with a specific environment for uploading documents to support RAG. This environment allows document bases to be specialized according to the nature of the request (e.g., marketing, technical, regulatory, etc.), in order to provide relevant and accurate responses.
GAIA also allows you to retrieve the original texts behind each response, so that you can check them immediately if necessary.

The modular nature of GAIA allows for modular implementation based on various Large Language Models (LMM).
The standard implementation uses ChatGPT, thus taking advantage of the continuous release of new features, as well as the immense computational power offered by full-web models.
Just consider that ChatGPT 4 is believed to be based on eight specialized models, each of which implements approximately 200 billion parameters (weights).
It is also possible to implement GAIA in “full edge” mode, i.e., supported exclusively by the resources of a PLC or industrial PC, without access to the web.
This architecture is based on language models from the LLaMA (Large Language Model Meta AI) family, which can also run on machines without a dedicated Graphics Processing Unit.
For reference, an LLaMA model with one billion weights occupies local memory in the order of 1 GB.
In addition, it is necessary to provide the space required to implement RAG locally.
This means that a full edge implementation of GAIA must be focused on a specific mission in order to achieve power and effectiveness comparable to the web version (e.g., technical support for resolving machine downtime, including technical documentation of the machine and a historical database of maintenance tickets in the RAG).
However, the modular nature of GAIA allows cloud and full edge chatbots to be used side by side, depending on the type of query and the need to be usable offline.

The G-Mation platform is particularly well suited to supporting the implementation of GAIA in both cloud and full edge modes.
First of all, the docker architecture allows you to easily implement and update GAIA’s dockerized apps, even on machines already in the field.
In addition, G-Mation’s webserver architecture, thanks to NGINX technology, allows you to manage additional backends, in addition to the one integrated in the PLC for machine control and configuration.
In particular, the backend contained in the GAIA docker allows you to call up the application directly on the G-Mation operator panel, based on web technology.
It is as simple as opening a new window in your personal computer’s browser.
As for cloud-based implementation, the availability of a dockerized VPN directly on the G-Mation PLC allows you to access the GAIA cloud platform in complete security, without having to go through additional gateways.
As for full edge implementation, the 8 GB of memory and the multicore architecture of the CPU allow you to perform many operator support tasks without affecting machine control in any way.