telexed ~ c / 9771e3a2-ec1radar:50 · otherLIVE
← back
NO.
#9771e3a2
Topic
OTHER
Source
GitHub Trending Weekly
Published
2026-05-11 02:09:49
Importance
★ 5/10 — radar 50
`PageIndex`: vectorless RAG with tree-based reasoning
FIG-9771:1

`PageIndex`: vectorless RAG with tree-based reasoning

Replaces embeddings and chunking with a document tree plus LLM-guided search. Interesting if vector recall keeps missing long-doc answers, but cost/latency need real-world validation.

[ KEY POINTS ]
  1. Retrieval is split into two steps: build a table-of-contents-style tree, then run reasoning-based tree search over it.
  2. The pitch is no vector DB and no chunking, with page/section references that make retrieval paths easier to inspect.
  3. It claims 98.7% on FinanceBench, so the strongest fit is document-heavy domains where relevance beats semantic similarity.
  4. Recent additions push it beyond a single file: file-system tree indexing, a chat product, and MCP/API access for agent workflows.
Originalgithub.com/VectifyAI/PageIndexRead original →

// related