From Unstructured Support Emails to a Productive RAG Knowledge Base

As gPlug products became more widespread, the volume of support requests increased significantly. Many requests were repetitive or referred to information already available in previous emails, documentation, or on the website.

At the same time, this knowledge was stored unstructured in email inboxes and could only be accessed with considerable manual effort. This led to inefficient processes and longer response times for support.

Data protection posed another key challenge. Processing data outside of Switzerland is prohibited. Requirements such as full data sovereignty, local data storage, and avoiding dependencies on international cloud providers had to be strictly adhered to.

We developed an AI-based solution using Retrieval Augmented Generation (RAG) that is operated entirely in Switzerland.

At its core is a multi-stage pipeline for the automated processing of historical support emails:

• Extraction of emails from the inbox

• Processing by a Large Language Model (LLM)

• Removal of personal data and irrelevant content (e.g., signatures)

• Structuring each email into question-answer pairs

• Generation of concise knowledge snippets

The processed content is then stored in a vector database and made available via the open-source chat interface Open WebUI. In addition, content from the technical documentation and the website is fed directly into the knowledge base.

We provided various preconfigured AI agents for use, including:

• a Q&A assistant for quickly answering support questions

• an email agent that generates draft responses to incoming inquiries

All model inference is performed using open-source LLMs hosted by a Swiss provider, ensuring that all data protection requirements are consistently met.

By implementing this solution, gPlug can, for the first time, make the knowledge contained in emails available in a centralized and structured manner.

Specifically, this results in the following benefits:

• Significantly reduced manual effort in support

• Faster and more consistent response generation

• Reuse of existing knowledge across all channels

• High-quality responses thanks to context-based AI support

• Full data sovereignty thanks to hosting and processing in Switzerland

• Compliance with all data protection requirements and complete independence from the U.S. Cloud Act

Support staff can now focus more on complex issues, while standard inquiries are handled efficiently and semi-automatically.