SWIRL vs Off-the-Shelf RAGs: A Study

We measured what six different RAG setups actually send the model to answer one question, and checked it against the frameworks' own source code. The honest headline first: SWIRL doesn't win by using smaller chunks. Anyone can lower top_k. On a lean query, LlamaIndex's default sends fewer raw tokens than we do. What SWIRL does that off-the-shelf stacks don't... for starters, it never sends the model duplicate or superseded versions of a document. In a versioned corpus, that's the whole ballgame, because top-k retrieval returns near-identical copies by design and every framework ships without de-duplication. And it answers in one bounded call, instead of the refine and tree-summarize modes that call the model once per chunk and multiply both cost and latency. The result is an answer grounded in the document your team approved, and a cost per query that stays flat as your corpus grows. Full study, with the caveats, in the comments!