‹ Case study

How COMAN brought its product knowledge together

One locally hosted assistant brings help articles, product documentation and answers from support emails directly into COMAN’s software.

Client
COMAN Software GmbH
Service
AI knowledge base, local AI, software integration
Year
2025–2026

Project report by Martin Kogut, AI Solution Architect at hypescale.

When someone logs progress in COMAN Mobile or sets up a user role in COMAN Manager, they usually need help with one specific task. COMAN Software GmbH wanted answers available inside its project management software. The company already ran a HubSpot chatbot. Its ongoing costs helped prompt the move to COMAN’s own infrastructure.

From help articles to a shared knowledge base

We started with the help articles COMAN already maintained in HubSpot. We indexed them for search through the existing integration, so the team could keep editing them in HubSpot.

We added more product documentation as the project grew. We reviewed support emails, grouped similar questions and wrote answers to the recurring ones. We added those answers to the knowledge base as a dedicated Q&A collection.

The assistant uses retrieval-augmented generation (RAG). It finds relevant passages in the knowledge base and passes them to a language model to write the answer. For COMAN we combined semantic search with keyword search. That helps users find answers whether they describe a task in their own words or use a precise name like COMAN Mobile.

Running on the DGX Spark

Ollama runs the language model on an NVIDIA DGX Spark. Qdrant provides the search index for the knowledge base. The application runs in Docker.

NVIDIA DGX Spark with a hypescale sticker beside its box on a desk

Setting up the NVIDIA DGX Spark for the COMAN project.

Testing the assistant with 370 product questions

For our internal evaluation on March 9, 2026, we ran 370 questions through both search and the full chat. Topics included access to COMAN Mobile, user roles and differences between license types. The project records list the sources retrieved and the answers generated separately.

When asked whether a license could be shared with a colleague, search returned four sources and scored them as weak matches. Questions like this depend on retrieving the right passage. In operation, we monitored response times and answer quality and reviewed individual cases where answers were slow or sources were a poor match.

What changed for COMAN

COMAN replaced its HubSpot chatbot. Ongoing costs are now lower and easier to forecast. The support team has one place to find product knowledge. For a new request, it can find an answer prepared for a similar case and the relevant passage in the documentation. Answers from individual support cases can now help with future questions.

Our AI knowledge base page explains how we define sources, access rights and answer quality before we build.

For more project details, see Martin Kogut’s portfolio. To explore the software, visit the COMAN Manager product page.

What would you simplify?

Bring a task that takes up time in your day. In the first conversation, we’ll look together at what could change.