Vectorless, reasoning-based RAG
Similarity is not relevance.
Relevance requires reasoning.
Vector RAG retrieves by semantic similarity — so it misses what is relevant but not similar, and returns what is similar but not relevant. PageIndex replaces the vector index with a hierarchical tree index and lets an LLM reason its way through it, the way a human expert turns to the right section of a long report. No vector DB, no chunking.
Two retrieval topologies, same question
“What was the 2023 operating margin?”Vector index
Chunks scattered in embedding space. Nearest neighbours come back; the route does not.
Result: opaque. The project calls it “vibe retrieval” — three chunks surfaced, with nothing to point at for why.
Tree index
A tree per document. An LLM walks it, and the walk itself is the citation.
Result: traceable to explicit references. The lit path is the answer’s provenance, not a post-hoc explanation. Node labels are illustrative.
The difference is a shape, not a setting. One topology can show you the route it took; the other can only show you where it landed.
Compared with vector RAG
Figures are the project’s published local-indexing benchmark, reproduced as stated and not independently verified here. Indexing is paid once; the project notes every later question reuses the tree.