How it works

Architecture diagram

Three configurations

All three use the same documentation snapshot, questions, answer model, prompt and output limit, and pass the top 5 passages to the answer model.

How it is measured

Evidence is marked as character ranges in the source documents, so the comparison is fair even though the configurations cut the text differently. Retrieval is scored without any LLM. Answers are scored by a separate judge model, validated against human labels on a sample.

Trace showing retrieval, fusion, reranking and generation

Source content: Kubernetes Documentation, © The Kubernetes Authors, licensed under CC BY 4.0. Normalized and chunked (modified). Full methodology, limitations and per-question results are in the GitHub repository.