Source: Own Illustration
In our last post, we introduced you to Ritchy, the IT-ServiceDesk’s AI-powered support chatbot. If you’re not yet familiar with Ritchy, we recommend reading the first post first. There you can learn more about the background and development of the chatbot.
When you select the support chat function on our IT Center website, in IT Center Help, or directly in RWTHmoodle, you can now choose whether you want to chat with Ritchy or with the IT-ServiceDesk staff. But if you’ve been wondering how Ritchy works in detail, we have the answer for you in this post.
How Ritchy Accesses Knowledge
The chatbot Ritchy is based on the latest language models from OpenAI (GPT5-CHAT), which are integrated via Microsoft Azure. In addition, Ritchy uses the Retrieval Augmented Generation (RAG) method to provide accurate answers. This is also provided by Microsoft Azure. Ritchy accesses the stored IT Center Help knowledge database to answer queries as accurately and contextually as possible.
RAG is an approach in which generative AI models not only draw on their trained knowledge, but also retrieve specific external information from connected knowledge sources. Instead of relying solely on the data in the language model, RAG’s AI first searches for relevant content, for example in a document database, and then uses this information to generate an informed response.
The big advantage here is that RAG significantly increases topicality and content accuracy, especially for context-specific or complex technical topics that are not included in the training data set. One example is queries about multi-factor authentication and the associated tokens, whose meaning at RWTH differs from the usual standard definition. RAG thus brings together the best of both worlds: precise information from the stored content of public documentation and natural-sounding language and logical structure from the implemented language model. This improves the quality of responses by basing support requests on verified and up-to-date knowledge.
Access to previously learned knowledge from the language models has been deliberately restricted in Ritchy. As a result, Ritchy only draws on its stored knowledge and does not actively view content from the internet. Some queries, especially those relating to content that is not available in the public documentation, may therefore not be answered.
In addition, a so-called system prompt has been stored, which determines how Ritchy should behave when responding to requests. Ritchy is configured to take on the role of a support employee when responding to requests. For each request submitted, this system prompt is passed on to the language model as context. The focus is clearly on providing support for IT problems.
If use cases go beyond the IT support context but still require the functionalities of an AI chatbot, this can be easily solved using RWTHgpt. You can read about the best way to formulate questions in RWTHgpt in the blog post “Ask Better, Answer Better.” There you will also find further information and background knowledge about RWTHgpt.
Working With Ritchy
AI systems reliably recognize patterns, but they do not know the user’s initial situation and find it difficult to understand it without the right information.
In order for Ritchy to provide the best possible answers, it is important to give the chatbot enough information when making inquiries. Similar to human support staff, context is needed to understand a problem.
A short statement such as “I forgot my password” is too general in this case. Additional information, such as which IT Center Service the password was forgotten for, is helpful. The following aspects should be taken into account:
- Concretization: The more precise the request or task in the request, the more accurately Ritchy can respond.
- Contextualization: The more context Ritchy receives for a request, the easier it is to assign it to existing topics. For example, when requesting a password change, always specify the associated service.
- Clarity: Try to avoid words that are used in multiple contexts with different definitions. For terms that are used in multiple subject areas in the RWTH context, you should specify the relevant topic. For example, if you are referring to tokens, it is necessary to specify whether it concerns MFA or RWTHgpt. And if it concerns MFA, you should specify exactly which service it is.
- Refinement: Prompting is not a one-way street. Especially in the current test phase, it is possible that Ritchy will not provide a complete answer right away or will need additional data. It is worth asking follow-up questions or asking Ritchy the question again with additional information.
If you still need more information at the end of a chat or want to communicate with IT-ServiceDesk staff, this is of course also possible. By selecting “Contact IT-ServiceDesk” at the top right of the web interface, you can decide whether you want to chat directly with support staff during service hours or whether you want to create a ticket in the IT-ServiceDesk system instead. You can also specify whether or not you want the previous chat history to be transferred.
Have you already used Ritchy and gained experience with it? Let us know in the comments.
Responsible for the content of this article is Robin Jakobitz.
Leave a Reply