Success Story: AI Support Chatbots in Intralogistics with Flexus

Background: Generative AI offers potential – but what is the best way to get started?

Flexus AG is a Würzburg-based company specializing in the optimization of intralogistics processes in SAP. With its 360° Logistics Suite, Flexus provides customers with SAP add-ons designed to streamline and enhance material flows. The technical team handles inquiries about the solutions and supports employees in resolving issues. All essential information about the software is documented in a comprehensive wiki, which serves as a central knowledge hub for the team.

Flexus came up with the idea of making life even easier for their employees and technical team by introducing an AI-powered chatbot. While Flexus already has highly skilled software developers with strong technical expertise, they have not yet gained in-depth experience with generative AI. They therefore turned to prognostica for a hands-on kickstart into the world of generative AI, enabling them to subsequently dive into development on their own. A key requirement for them was the use of open-source LLMs, allowing the models to be hosted internally and giving them full control over model versions, data, and computing costs.

“We see generative AI as an opportunity to make our software solutions even more user-friendly and innovative. We just didn’t know where to start.” – Flexus AG

Solution

“There are many companies that already call sophisticated Excel spreadsheets AI. What we’ve achieved with prognostica truly deserves the AI label.” – Flexus AG

Approach

It was important for us to use an agile approach to quickly arrive at a first prototype that made artificial intelligence tangible for Flexus. After all, when working with real-world data, not everything can be precisely planned in advance. Through close discussions with Flexus, a shared understanding emerged, along with concrete ideas of where the journey could lead.

The first prototype chatbot, WIKInger, was up and running within just a few days: it enabled Flexus to query its own internal wiki. The chatbot was built on a RAG pipeline (Retrieval Augmented Generation), in which a suitable open-source LLM was enriched with company-specific context to answer questions about Flexus’ software documentation. Flexus could now test the chatbot, share feedback with us, and gain first-hand experience. Step by step, we worked together to solve problematic cases, shape the chatbot’s individual tone of voice, and implement additional requirements. For example, the chatbot was designed to understand Flexus’ specific technical vocabulary and to handle information stored in tables.

Many factors come into play when it comes to turning a prototype into a stable application. We advised Flexus on how to keep data, context, and the chatbot synchronized so that it can always access the most up-to-date information in the wiki. In addition, the chatbot needed to be able to collect user feedback on its responses in order to continuously improve the relevance and quality of its answers. With our support, Flexus also gained the ability to update the underlying LLM models at any time and replace them with alternatives if needed. Role-based security was another key topic: how to ensure that only authorized personnel can access certain information?

In order to host such chatbots in-house in operational use, Flexus also needed a suitable and stable MLOps and IT framework as a foundation. We therefore supported the Flexus developers in selecting an appropriate graphics card, setting up the necessary pipelines, and getting the complete framework up and running – including the Kubernetes cluster, authentication mechanisms, and the user frontend.

Outcome: Sustainable Enablement for Long-Term AI Success

With WIKInger, Flexus now has an LLM-based chatbot that allows their technical team to respond to employee inquiries even more efficiently – faster, more accurately, and with source references for the information provided. The solution also serves as a blueprint for additional LLM-based applications that Flexus can develop independently in the future.

By using open-source LLMs, Flexus remains independent of any single provider in the generative AI space and maintains full control over model versions, data, and costs – especially when it comes to sensitive information.

“We now feel very comfortable navigating the world of LLMs and, thanks to prognostica, are capable of building chatbots and hosting LLMs ourselves.” – Flexus AG

The Project Team

Manuel

Senior Data Scientist

André

Data Scientist

Would you also like a hands-on quick start in LLMs and generative AI? Then feel free to