Cite sources. Keep a human in the loop where mistakes are expensive. Iterate. Most chatbot projects fail not because the technology is hard but because they are not grounded in real data and real user needs.
The clearest internal chatbot guide is a worked example. A leading member network came to us with an unreliable knowledge chatbot. Off-the-shelf OpenAI Assistants with file search had not delivered the accuracy they needed. The chatbot had to live inside HumHub – their existing social network – and be operated by a non-technical team.
We delivered Version 1 in four short sprints: system architecture, RAG implementation with semantic search, HumHub integration, full documentation. Total estimated effort 10 working days. Stack: n8n in Berlin for workflow orchestration, Qdrant in the EU for vector search, GPT-4 served via Microsoft EU sovereignty as the LLM. Optional fully open-source EU alternatives: Mistral Medium 3 as the model, Milvus as the vector database. Version 1 is now in production serving more than 1,000 network members, time-aware, operated by a non-developer team.
The same RAG architecture supports GDV (German Insurers Association) across tens of thousands of policy documents for 400+ member companies, and an AI Member Platform for a leading German association combining chat-based discovery with traditional category filters.
RAG, not fine-tuning
EU-compliant stack
Cited and verifiable
Production accuracy
A leading member network chose RAG after off-the-shelf OpenAI Assistants proved unreliable; ours runs in production for 1,000+ members
Time to value
production chatbot in four short sprints, the way the RAG chatbot for a leading member network was delivered
Maintained by your team
low-code architecture so non-developers can operate and extend it after handover
Hallucinations under control
answers grounded in your real documents, sources cited and clickable, human-in-the-loop where it matters
EU AI Act ready
risk classified, GDPR aligned, with audit trails and citations built into the architecture
1. Architecture and scope
2. Build and iterate
3. Hand over
We bring a decade of impact-tech experience and more than 160 AI projects since 2019. Through our free AI for Impact course, more than 100,000 people have learned how to use AI for the common good. We do not run inspiration days. We run scoping sessions and build engagements that ship, the way we have delivered AI for the organisations below:
Off-the-shelf assistants with file search are not enough for production. A leading member network tried this first. It was unreliable. RAG with proper grounding and citations is what makes the difference.
Fine-tuning is almost never the right answer. It is expensive, hard to maintain, and does not solve the hallucination problem. Use RAG against your real documents instead.
Hallucinations cannot be fully eliminated. Mitigate them with grounding, citations, restricted prompts, and human review on consequential outputs. Accept that 100 percent accuracy is not the goal; defensible accuracy is.
Should we fine-tune or use RAG?
Which model should we use?
How do we handle hallucinations?
How long does it take to ship?
Where does our data go?
Can our team maintain it after delivery?
Tell us about your documents and the questions your team or members keep asking. We will reply with a proposal and a date.
Simon Stegemann
Co-Founder and CEO
