How the Brändi-GPT enables barrier-free access to information

We usually use artificial intelligence and chatbots as a matter of course, for example to obtain information. However, access to information and searching on public websites is often not barrier-free for people with disabilities. Internal documents are also often difficult to find. The Brändi Foundation is now using its own chatbot, which uses an inclusive search function to make both public and internal documents quickly and easily accessible.

Knowledge management is a major challenge in all companies. Agile procedures, modern organisational structures and an ever faster pace of change make this task even more difficult. They increase the need for company information to be adapted and found quickly.

Access to company information is particularly difficult for people with disabilities, as internet or intranet sites are not always designed to be barrier-free from the ground up. Even if they are, this does not automatically mean that information is easy to find. A full-text search, for example, places demands on the user’s comprehension and formulation skills. Searches with text entries that contain spelling errors often lead to no results.

Brändi addresses this problem with the Brändi GPT. The solution makes it possible to find information from the company quickly and easily. The search is designed to be inclusive and also supports people with disabilities. This is achieved through the targeted use of AI.

Using AI for inclusion

The solution used combines the strengths of query-based techniques with generative AI models. Query-based models are excellent at extracting information from existing sources such as SharePoint pages or any documents. However, they cannot provide customised answers. Generative models, on the other hand, can generate specific, contextualised answers, but their accuracy can sometimes be problematic (hallucination effect). To overcome these weaknesses, the Retrieval Augmented Generation (RAG) approach is used, which combines both techniques from AI.

The aiaibot platform from swiss Moonshot AG uses its own modules for the query-based technique as well as the OpenAI service from Microsoft CH, which is based on the LLM from OpenAI.

Firstly, the relevant documents are copied into a separate environment (knowledge base) that is only available to Brändi. A decision can be made at document level as to which information is accessible. Documents that should not be accessible to all employees for internal confidentiality reasons are not copied to the knowledge base.

Users of the Brändi GPT can formulate their questions via a prompt. The aiaibot platform uses a semantic search to identify relevant documents or suitable text sections from the knowledge base. These are then forwarded to the LLM (currently GPT-3.5 or GPT-4.0) via the OpenAI Service API. Based on the relevant knowledge, the LLM generates a suitable response, which the aiaibot platform outputs to the person making the enquiry.

The user interface for the Brändi GPT can be integrated into SharePoint pages or applications as required.

Figure 1: Internet start page from the Brändi test environment

LLM can also be used to interpret queries that are not formulated correctly or use related terms. For example, if you search for the event “spring magic” although it is labelled “spring magic” on the website, the event can still be found. This represents a new way of obtaining information that was previously not possible. Weak knowledge of German or difficulties with wording are no longer barriers to this inclusive access to information.

The content of PDF documents can also be interpreted, such as the weekly menu plan from the restaurants. If the menu plan is stored as a PDF document, it is possible to ask specifically for gluten-free dishes, for example. Such a query would not be possible in a simple text search on an intranet page. In addition to the answer, the document on the basis of which the answer was generated is also referenced.

Data protection

The Brändi GPT can be used in various configurations: for visitors to the website, all Brändi employees on the intranet pages or for specific applications at Brändi. A clearly defined and delimited knowledge base is used for each application group.

This knowledge base is hosted in Switzerland. Microsoft Azure Switzerland guarantees that neither the prompts nor the generated answers will be made accessible to other customers or OpenAI. They will also not be used for further training purposes.

Realisation

In the first phase, three use cases will be implemented at Brändi:

• Brändi GPT on the website: This GPT is specifically geared towards information on the Brändi website and is offered to all visitors to the website. It can provide answers based on information from all Brändi websites. This GPT does not recognise general knowledge information from the world or internal Brändi documents and therefore does not answer any questions in this regard.
• Brändi GPT on the intranet site: This GPT should have access to the internal documents from the Brändi SharePoint. All Brändi employees should be able to search for company information, regulations and directives here.
• Brändi GPT in the IT ticket system (JIRA): This GPT should have access to several thousand IT support tickets. This will allow IT support employees to search for relevant information directly in the complete knowledge database.

Further use cases are planned for the future:

• Integration into the web shop
• Support when writing reports or entries in applications
• Finding information from previous projects

In summary, the result is inclusive access to all desired company information for people with and without disabilities. This new type of knowledge management can significantly improve the efficiency of work processes.

More about the Brändi Foundation

The Brändi Foundation is a foundation under private law and a professional non-profit organisation based in the canton of Lucerne that promotes the professional, social and cultural integration of people with predominantly mental or psychological disabilities. To this end, it runs a total of 15 of its own companies, which are active in 14 sectors in the areas of production and services and work closely with industry, commerce and the canton of Lucerne. With a total of over 2000 employees, it is one of the largest employers in Central Switzerland.

The Brändi Foundation is a partner of the Institute of Digital Technology Management (IDTM). In addition to his work at the IDTM, Prof. Dr Andreas Liedtke is also Head of Digitalisation and ICT at the Brändi Foundation, where he oversaw the Brändi-GPT project.