Mimiron AI

Playground

Ask a question about Kubernetes. The system finds 5 passages in the Kubernetes manual, then an AI writes an answer using only those passages. Change the settings and see what changes. Your own question has no score, so the box below shows how the settings you picked did on our test questions.

The manual here is 31 pages of the Kubernetes documentation, not all of it: pods, health checks (probes), networking and services, ConfigMaps and Secrets, memory and CPU. Questions outside these pages get no answer. Try one:

Chunking strategy

RecursiveCharacterTextSplitter: splits on \n\n, then \n, then sentences, until a chunk fits the size limit.

Retrieval mode

Vector and BM25 run together; scores are normalised and fused at alpha 0.5 (relative score fusion).

Embedding model

snowflake-arctic-embed-m-v1.5, 768 dimensions. The model the evaluation scored.

Reranker

Cross-encoder gte-reranker-modernbert-base scores each (question, chunk) pair and reorders the top 5.

Hit rate@5 on the eval set: 101 of 106 answerable questions. This is the winning setup.