AI has sparked an unprecedented global growth race. Still, quality (and truth) often lacks in the answers from our AI-tools. There is a way to fix this, though! A solution which gives your AI a chance to actually understand you – and thereby think and answer like you. It just needs to access just about everything. But is that safe?
Roles and relations
Would you grant a random AI access to your documents and data? Short answer, NO. But, if you had your own AI, in your own cloud-solution – placed in your own datacenter, then it would be a different case. We will save the security aspect a bit and start out by focusing on the solution itself: Some people call it Context-aware generation. Others call it Grounded AI. The technical term itself is quite a mouthful, Retrieval-Augmented Generation. For everyone’s best we call it a RAG-solution.
Simplified, RAG is about letting your AI make its conclusions from your data, your documents, texts and strategies. It needs access. You must show it where to search – and tell it how to answer. Instead of basing its answers in historical, generic data from everywhere in the World, then your AI will now “guess” less and instead base its answers on (your) facts.
AI-specialist and Director, Innovation & Emerging Technologies, Leif Elgaard Høj, says: “With a RAG-solution in your system you go from general intelligence to narrow intelligence. If you want a GPT-like solution in your organization, then you need it to answer correctly. However, the problem is that AI’s make up things. We are close to the truth, but we do not know for sure if the answers are true”, he states and adds: “We can work around this, though. We can tell our AI how to behave. Not only what it should be answering, but how to answer. Do we have something in our system on shipping-taxes? You must answer like if you were from the Tax-authorities. By doing this we give our AI roles and relations”.
RAG-solutions, in short
Retrieval-Augmented Generation is a technology which enhances the quality in the use of AI. This is done by granting the AI access to data, from which it then retrieves its answers. That way the AI will “guess” less and instead answer from facts. Another advantage with RAG is that new knowledge and data can be added without having to re-train the AI’s core-model.
Protect your security and business with your own, private cloud
Leif Elgaard Høj answers directly on the question of sensitivity of data and security in RAG-solutions. He explains that AI and data are inextricably linked, but AI are also inextricably linked to cloud-solutions. “From a starting point, when using an AI, you upload your material to a public cloud-solution. There will always be a risk related to that. Data can be exposed – you do not have full control. In Danoffice IT, this is a scenario, we always wish to avoid. Therefore, we offer our clients a Private Cloud for AI-solution, a PCAI. Here, all the client’s data are secure and stored on-premises and with full control”.
What is Private Cloud for AI?
Most are familiar with the concept of a cloud-solution: We have some data which we save in what we just call the cloud. Where our data are, however, we do not really know, but we can access them, and it is efficient. This the characteristics of a so-called, public cloud. It is easy, but security is somewhat “up in the air”. No joke. Add to that the challenges of controlling the costs.
The closest we get to the opposite of a public cloud, is a so-called on-premises datacenter. A local, physical datacenter. Until recently some would label this “the old model”, but due to the new World-order, the old on-premises-solution is as hot as ever – due to concerns over our data. The reason: With an on-prem-solution we can build a private cloud-solution. With this your organization will have all the benefits known from a public cloud, but now the cloud is yours alone.
The newest addition to that is AI. And since AI builds on cloud-solutions then the approach is the same: If you use public AI in your operation, then you are vulnerable. Therefore, we are offering the strongest possible Private Cloud for AI-solution to our clients.
Control and context
”If you, as a district court, a municipality, or another public organization need a fast discission on a complex matter, then you can use AI and save resources. However, you need a system which answers in a very specific way and that knows the difference between many different types of cases. And importantly, it needs to know all imaginable legislation at hand, for example on sensitive personal matters. If you were to use a public GPT, then the answers would differ from each other, and information cannot be recreated”, Leif says before elaborating on a critical aspect: “When we are dealing with personal cases or matters there needs to be governance, GDPR and heavy data-compliance. All-in-all solutions like this should only belong on-premises – meaning in Private Cloud-Solutions. This is our only way of giving the AI ultimate context while still being in full control”.
Have a nice journey
In Danoffice IT we have created many RAG-solutions. A recent example is a large, complex system to public organizations, where data integrity needs to be extreme. A RAG-solution where we have configured the AI-agents in a hierarchy, as Leif explains. “This system is built in layers. In the top-layer we have governance and regulation for behavior, access, and compliance. After that we have the retrieval-layer, where the AI finds relevant information in the organization’s own sources. Next is the orchestration-layer which controls how the tasks are solved. And finally, in the bottom-layer we have the actual answering-layer where tone-of-voice, formality and professional discourse is shaped. This gives the AI-solution a much higher level of control and consistency”, he states. “We have built a system with a hierarchic relation for your AI-agents, which ensures full control and compliance throughout your AI”
All this is only possible if you assign your RAG on a PCAI – a Private Cloud for AI. And the advantages are many, says Leif. “With a PCAI we are not only in control of security. Unlike a public AI we do not need to train our AI on its competencies. This is an important factor, since this takes heavy resources. Thus, it is a bonus for having an AI-journey with maximum agility and room for development. By freeing up resources from training and maintenance we can dedicate those to other tasks, which creates value in our operations”, concludes Leif Elgaard Høj, Director, Innovation & Emerging Technologies in Danoffice IT.
