PageIndex: Vectorless RAG Where an LLM Navigates a Document Tree

1. What It Is: Stars, License and Scope

PageIndex is a Python SDK from Vectify AI. Its README calls it a vectorless, reasoning-based RAG engine: each document gets a tree index (think table of contents with summaries) and the chat model searches that tree instead of running a similarity query over chunk embeddings.

2. The Architecture in 4 Steps

Putting together the README, the PageIndex Flash post (2026-08-26), the original PageIndex post (2025-09-19) and the package’s agent tools, the flow is four steps.

3. Quickstart (Verified Against the README, 2026-10-14)

The install command and both snippets are copied from the README. I did not execute them: the pip install and the LLM calls are yours to run in a clean virtual environment with your own key.

import os
from pageindex import PageIndexClient

os.environ["OPENAI_API_KEY"] = "your-openai-key"

client = PageIndexClient(
    index="gpt-6-luna",                 # model to build the tree index
    chat="gpt-6-sol",                   # model to search the tree
)
doc_id = client.submit_document("report.pdf")["doc_id"]

answer = client.chat("What was the 2023 operating margin?", doc_id=doc_id)
print(answer)

import os
from pageindex import PageIndexClient

os.environ["PAGEINDEX_API_KEY"] = "your-pageindex-key"
os.environ["OPENAI_API_KEY"] = "your-openai-key"

client = PageIndexClient(
    index="cloud",                       # build and store the index in PageIndex Cloud
    chat="gpt-6-sol",                    # use your preferred compatible model for chat
)
doc_id = client.submit_document("report.pdf", wait=True)["doc_id"]
print(client.chat("What was the 2023 operating margin?", doc_id=doc_id))

4. Local vs Cloud: What the Open-Source Mode Does Not Do

This is the part most summaries skip. The README’s own table, with the Cloud column being the vendor’s hosted service:

The practical consequence: a pile of scanned contracts, or slide decks where the meaning sits in figures, is a Cloud job (OCR and image understanding run there). The local package reads only the PDF’s text layer.

5. Cost and Latency: What LLM Tree Search Really Costs

A vector query is one embedding call, an index lookup and one generation. A PageIndex query is a multi-step LLM loop: the answering model reads the outline, requests pages, judges whether it has enough and may go around again. You pay for that in tokens and in seconds. These are the vendor’s own numbers, all self-reported.

6. Where It Fits

These are the problem shapes the design targets, based on how the vendor describes it. I have not tested them myself.

7. When NOT to Use It

This is the section I would have wanted before adopting it.

Conclusion

PageIndex is a clear, well-scoped idea: let the model read the outline the way a person would, then open only the pages it needs. The open-source package gives you that for text-based PDFs on your own machine, with an LLM bill that is small at indexing time and larger at query time.

What it does not give you for free is OCR, folders, metadata, a hosted MCP server or the multi-document tree layer. Those are the vendor’s cloud and enterprise offering, and it is worth knowing where that line sits before you design around it.

Sources

• VectifyAI/PageIndex: README, LICENSE and repository (MIT, Python)
• PageIndex v0.4.0 release notes (2026-10-12)
• PyPI: pageindex 0.4.0 (Python 3.10+, MIT)
• PageIndex docs: Getting started (install, local vs cloud, index= and chat=)
• PageIndex docs: Agents (agent_tools, as_openai_tools, citations)
• PageIndex docs: MCP server (API key, api.pageindex.ai/mcp)
• PageIndex blog: Next-Generation Vectorless, Reasoning-based RAG (2025-09-19)
• PageIndex blog: PageIndex Flash (2026-08-26)
• PageIndex blog: Scale PageIndex to Millions of Documents (2026-05-03)
• VectifyAI/PageIndex-OSS-Benchmark: 62 questions, 34 PDFs, results table
• PageIndex blog: Mafin 2.5 and the FinanceBench claim (2025-02-19, vendor-reported)