Best Practices for Building an AI Development Workflow with Qlik Answers

[TarunNagar]

Hi everyone,

I'm currently exploring AI development using Qlik Answers and would like to build an enterprise AI assistant that can answer questions from both structured and unstructured business data.

I have a few questions:

What are the best practices for organizing knowledge bases for accurate AI responses?
How do you minimize hallucinations when using enterprise documents?
Are there any recommended approaches for integrating Qlik Answers with external LLMs or AI applications?
How do you handle security and user permissions when exposing business data to AI assistants?
Any real-world examples or lessons learned would be greatly appreciated.
Thanks in advance!

[LaurentNotes]

A good starting point is to keep the knowledge base focused rather than feeding every document into one large collection. Separate content by business area and make sure documents have clear ownership and update dates. That makes retrieval more reliable and gives you an easier way to remove outdated information.

For hallucinations, grounding the answers in retrieved source content seems more useful than relying on the model alone. I would also return the source document or section with each answer so users have a way to verify the response.

Security is probably the part worth planning early. The assistant should respect the same permissions users already have for the underlying Qlik data and documents. Otherwise, even an accurate answer becomes a data access problem.

I'd be interested to hear how others are handling document updates. Keeping the AI knowledge base in sync with changing business documents seems like one of the harder parts in a real deployment.

[hugo_andrade]

@TarunNagar ,

Keep each knowledge base narrow, one domain, one owner. Qlik Answers ranks on semantic relevance, not on the most recent, so an old policy will compete with the current one and sometimes win. Retiring content matters as much as adding it. Against hallucinations, the control that holds is a regression set of 30 to 50 questions with known answers, re-run after every content change. With the valid content added to the knowledge base, hallucinations are very rare.

On access, permissions sit at the space and assistant level, not per document. Source-system file permissions do not follow content into the index, so anyone who can use the assistant can reach anything behind it. Segment by audience and test as a low-privilege user before go-live.