RAG-Wissen — Production RAG for our own knowledge base
RAG-Wissen is our internal RAG system for advisory knowledge, technical documentation, and market intelligence. Runs in production with Qdrant v1.17.1 and Harrier embeddings, provides the research base for our blog articles, and connects via an MCP server to Claude Code, Mingly, and our internal tools. Single source of truth for every advisory sparring session.
Embedding model choice — five models benchmarked
We benchmarked five embedding models for German-language content: mxbai-embed-large, Qwen3-Embedding, Jina-v3, Cohere-Multilingual, and Harrier. Harrier won with roughly 50 percent better NDCG@10 over mxbai and roughly 12 percent over Qwen3 — measured on a manually annotated test set from our own corpus. Result: production use since April 2026. The lesson: default recommendations from English-language benchmarks (often MS MARCO, BEIR) do not translate one-to-one to German advisory content with domain vocabulary.
Aggressive reduction — why we deleted 949 lines
RAG systems tend to accumulate piles of features over time: hybrid search, cross-encoder reranking, query expansion, multi-hop, reranking the reranker. In one sprint we removed 949 lines from RAG-Wissen — features we had built without delivering measurable lift in A/B tests. Result: less code, faster responses, lower cloud costs, simpler maintenance. This experience is valuable because it contradicts the widespread assumption that more RAG components always mean better retrieval. The blog post is linked below.
Practice proof for RAG strategy workshops
From RAG-Wissen we bring three things into every RAG strategy workshop: a benchmarking protocol you can adapt to your own content; a list of RAG features whose effort almost never pays off in practice; and an empirically grounded answer to the most common question — 'how do I measure whether my RAG system actually got better'. Plus the concrete architecture of a production system used daily.