PageIndex — Vectorless, Reasoning-Based RAG

One sentence: instead of embedding chunks and ranking by cosine similarity, PageIndex builds a table-of-contents-style tree for each document and lets an LLM walk that tree — the way a human expert turns to the right chapter — then answers with references to the exact nodes it read.

Mechanism (two steps)

1. Index — generate a hierarchical tree for each document. Upstream says the tree structure is extracted from the document layout without an LLM; the index model only summarises/refines nodes, so a basic model is enough.
2. Retrieve — an LLM agentically searches that tree. No vector DB, no chunking, retrieval sees conversation history and domain context, not just a query embedding.

	Vector RAG	PageIndex
Index	vector index over chunks	tree index over document structure
Retrieval	semantic similarity	LLM reasoning over the tree
Result	opaque ranking	traceable to explicit page/node references
Context	query embedding only	full conversation + domain context

Quickstart (local mode, from the upstream README)

pip install -U pageindex

import os
from pageindex import PageIndexClient

os.environ["OPENAI_API_KEY"] = "<your key>"        # never commit; use env/secret store
client = PageIndexClient(
    index="gpt-5.6-luna",   # model that builds the tree  — a basic model is sufficient
    chat="gpt-5.6-sol",     # model that searches the tree — use the best you can afford
)
doc_id = client.submit_document("report.pdf")["doc_id"]
print(client.chat("What was the 2023 operating margin?", doc_id=doc_id))

Model names above are copied from the upstream README as of 2026-09-30 — confirm they exist on your provider before running; substitute any compatible model.

Local vs Cloud

	Local (open source)	Cloud
Handles	text-based PDFs	text-based, scanned, image-rich
Indexing	on your machine	managed
Storage	local directory	cloud
Citations	page-level	block-level
OCR / image understanding	no	yes
MCP server	no	yes (needs a PageIndex API key)

Consequence for this repo: scanned regulatory PDFs (typical for tender documents, older HTA/market-access dossiers) are the weak case locally — OCR first (see ocrmypdf in the Batch 97 self-hosted list) or use Cloud.

Vendor-reported numbers (treat as claims, not facts)

• ~$0.001 per page to index locally with the index model named in the README (≈ a little over $1 for a 1,000-page textbook); 9–1,098-page benchmark PDFs indexed in roughly 13 s – 4.5 min.
• 98.7 % on FinanceBench (Mafin 2.5 write-up) vs ~50 % for a vector-RAG baseline. The comparison baseline is the vendor's; a well-tuned hybrid (BM25 + dense + reranker) may close part of that gap.
• Own benchmark: VectifyAI/PageIndex-OSS-Benchmark, 62 lookup questions — read before trusting for your domain.

When to use it vs the other RAG skills here

Situation	Use
One long structured document, need page-cited answers	pageindex
Large mixed corpus of short documents, need semantic recall	rag-implementation / rag-pipeline-architecture (+ hybrid-search-implementation)
Deciding build order for a RAG product	docs/procedures/rag-build-order.md
Millions of documents	PageIndex File System (Cloud-only, per upstream) — not covered by the OSS repo

Guardrails

• It sends document text to your chosen LLM provider. Do not index confidential, patient or client documents through a third-party API without a data-processing basis; use a local/on-prem model or the vendor's VPC option.
• Tree quality depends on layout: flat OCR text, slide decks and spreadsheets give weak trees.
• Always spot-check 5 answers against the cited pages before relying on a corpus (see /pageindex-ask).