Case Study: GoodData powers AI analytics with ~100ms semantic search on Qdrant

Key results

The challenge

GoodData's early AI-assistant prototype loaded its entire semantic model into the LLM context for every query, which drove high compute costs, slow response times, and hit model token limits. Most GoodData customers work with complex data models spanning tens or hundreds of datasets, making that approach unsustainable at enterprise scale.

The solution

GoodData moved to a Retrieval-Augmented Generation strategy backed by Qdrant's vector database, deployed on Kubernetes via Qdrant's official Helm chart after evaluating DuckDB and pgvector.

The results, in context

With Qdrant, embedding updates complete in seconds (hundreds to thousands per minute), semantic search results return within 100 milliseconds, and AI-assistant responses are maintained at approximately 5–10 seconds with negligible Qdrant latency overhead. GoodData serves over 140,000 end customers.

Products used