PageIndex Tree Retrieval
Status: stable
Overview
Section titled “Overview”Documents (PDF, markdown, text, legal) are indexed as a hierarchical table-of-contents tree. The calling agent navigates it with six pageindex_* MCP tools (or /v1/pageindex/*), reading only the pages it needs. No chunking, no embeddings, no LLM inside Xavier. Inspired by Vectify AI’s MIT-licensed PageIndex; reimplemented from scratch in Rust.
Evidence
Section titled “Evidence”On the public PageIndex-OSS-Benchmark (62 questions, 34 PDFs) an agent using the tools reached 61/62 reading 11.6 pages per question, and 60/62 reading 1.7 pages per question with in-document search. It complements BM25/vector retrieval and does not replace it.
Learn more
Section titled “Learn more”Full guide, rationale, attribution, evaluation tables and setup: docs/features/pageindex.md in the repository, and ADR-036.