KB Arena
Runs the same document corpus and question set through 19 retrieval architectures and reports citable, statistically tested evidence for which one fits a given dataset best.
KB Arena compares retrieval architectures on a user's own documentation so a team can choose one with evidence rather than guesswork. It runs the same corpus and question set through 19 built-in strategies — lexical, dense, graph, hybrid, hierarchical, reranked and more, including LightRAG, RAPTOR, PageIndex, BM25, HyDE, multi-query, SPLADE and two quantum-inspired rerankers — and records retrieval quality, answer quality, latency, cost and run artifacts for each.
A packaged demo with precomputed results needs no API key, Docker service or Neo4j instance; a guided /decide workflow then walks from picking a corpus and an optimization target through the comparison to a recorded decision that states the corpus, question count, metric, means, confidence interval and statistical-significance test, and explicitly notes what the comparison could not verify, such as an unchecked question set or top-k.
It installs with pip and includes a CLI.