I benchmarked tree-RAG against plain chunks. The boring one won.
Before shipping large-PDF indexing I had to answer a question properly rather than by reading opinions: how do you cut a document into pieces an AI can actually search? Fixed-size chunks against the tree-RAG approach everyone recommends, tested and measured. The boring method won, and this post publishes the result that argued against the more interesting architecture.
The boring method: fixed-size chunks
What it is, and why it is the default nobody writes conference talks about.
The clever method everyone recommends: tree-RAG
Hierarchical summarisation — the approach that sounds obviously better and gets recommended accordingly.
So I tested it — thoroughly
The benchmark setup, run against real documents rather than a toy corpus.
The results
Retrieval quality, and the token cost of getting it. Tree-RAG lost on the second while not clearly winning the first.
The honest caveat
Where tree-RAG would still be the right answer, and why the result is not a general law.