dalao.ai

PageIndex

Vectorless, reasoning-based retrieval for long documents, with a tree index and page-level citationsOpen source

About PageIndex

PageIndex is a long-document retrieval engine with no vector database and no chunking: it builds a tree index per document, then an LLM reasons along the tree to the relevant sections and answers with references. The open-source SDK has a local mode in which PageIndex Flash takes structure from the PDF layout, and a cloud mode that adds OCR and image understanding. As of 2026-10-26 the PyPI package pageindex is at 0.4.7, released on GitHub 2026-10-24 19:12 (UTC+8). The code is MIT-licensed; the cloud charges $0.01 per page indexed and $0.001 per active page per month, first 1,000 pages free.

EditorialUpdated 8 min read

What problem it solves

The common RAG approach cuts documents into fixed-size chunks and retrieves them by vector similarity. The PageIndex README argues that similarity is not relevance: in long professional documents such as financial reports, regulations and technical manuals, the passage that answers a question may share no wording with it, and finding it takes context and multi-step reasoning.

PageIndex works in two steps. Indexing builds a hierarchical tree for each document, with titles, page ranges, summaries and nested sections, in effect a table of contents for an LLM. Retrieval has the LLM reason down the tree with the query in hand, visit only the nodes on its path, and answer with references. No vector database and no chunking are involved.

Product shape. The open-source SDK ships as the PyPI package pageindex, needs Python 3.10 or later, and runs in a local mode and a cloud mode. Alongside it are the Chat App for the browser, PageIndex Cloud for developers, an MCP server, and an Enterprise edition deployed into a customer's own cloud.

The repository is VectifyAI/PageIndex, created 2025-04-01 18:53 (UTC+8), with main as its branch. Recent releases: v0.4.5 on 2026-10-23 00:40 (UTC+8), v0.4.6 on 2026-10-24 00:10 (UTC+8) and v0.4.7 on 2026-10-24 19:12 (UTC+8).

The publisher is Vectify AI. The contracting entity in the terms of service is Vectify AI Limited, company number 14827188, and Vectify AI is the copyright holder in LICENSE. The About page lists four founders: Mingtian Zhang, Yu Tang, Peter Hayes and David Barber, and describes the team as AI researchers with research track records.

similarity ≠ relevance — what retrieval actually needs is relevance, and relevance requires reasoning.

PageIndex README

Core capabilities

The tree is built without an LLM. The README states that the tree structure is extracted from the document layout without an LLM, and the index model only summarizes and refines it, so a basic model is enough. The chat model that searches the tree drives accuracy, and switching to a stronger model or raising reasoning effort needs no re-indexing.

PageIndex Flash. Released 2026-08-26, Flash is the default indexer in local mode for long text-based PDFs. It reads the layout the PDF already carries instead of asking a vision model to infer an outline from scratch. The blog puts indexing at about $0.001 per page with gpt-5.6-luna as the index model, and 13 seconds to 4.5 minutes for benchmark documents of 9 to 1,098 pages (as reported by the vendor).

Block-level citations. The changelog says that from 2026-09-16 cloud answers cite the exact layout block rather than the page, and only cloud documents indexed since 2026-09-13 carry blocks. The developer page says each block resolves to a bounding box that a custom viewer can highlight. Local citations stop at the page.

Multi-document search and integrations. The SDK supports streaming, multi-document search and citations, and the README lists the OpenAI Agents SDK and the Claude Agent SDK among agent frameworks it plugs into. Since v0.4.5 it can call Claude models through Claude Code on a Claude subscription, via the pageindex[claude] extra.

PageIndex File System. A 2026-05-03 blog post introduces a file-level tree indexing layer for reasoning over a whole document corpus rather than one document; the README marks it as cloud-only.

Who uses it

The developer page cites 35k+ GitHub stars and 30k+ developers on the cloud (as reported by the vendor). As of 2026-10-26 the GitHub API shows 39,685 stars, 3,430 forks and 16 contributors, and PyPI lists 53 releases of pageindex.

Benchmark results (as reported by the vendor):

  • On FinanceBench, Mafin 2.5, built on PageIndex, scores 98.7% accuracy.
  • On 63 questions over NVIDIA's 10-K filings, the developer page shows PageIndex at 100% accuracy for $110 per 1,000 queries, against vector retrieval with six embedding models at top-k 10 to 100, each model giving a range, best Voyage 4 at top-100 with 92.1% for $134 per 1,000 queries.
  • On text lookup questions from MMLongBench-Doc-V2, accuracy is 85.5%, 90.3% and 96.8% with gpt-5.6-luna, gpt-5.6-terra and gpt-5.6-sol as the chat model.
  • Against handing the whole PDF to the model, the native route costs 2.1 times more at 52 pages and 16.6 times more at 420 pages, and at 805 pages the document no longer fits the context window.

The pageindex-benchmarks repository publishes the data, code, every model call and the bills behind these results for recomputation.

How it compares

Against a general RAG framework: LlamaIndex is MIT-licensed, describes itself on GitHub as the document processing platform for AI, and offers many retrieval methods including vector indexes. PageIndex does only tree indexing with reasoning-based retrieval, and its docs state no vector DBs or chunking.

Against graph-based RAG: Microsoft GraphRAG is MIT-licensed and uses LLMs to extract entities and relations from unstructured data into a graph. PageIndex takes its tree from the document's own layout, the README states this extraction uses no LLM, and the structure is limited to one document's section hierarchy.

Against a code graph: Graphify, under Apache-2.0, builds a codebase and its surrounding docs into a knowledge graph for AI coding assistants. PageIndex targets long professional documents such as PDFs and does not model code structure.

The README's own table sets vector RAG beside PageIndex: the former uses a vector index with semantic similarity search, the latter a tree index with LLM reasoning, with results traceable to explicit references.

Who it's for / Not for

Good fit:

  • Teams working with long professional documents such as financial reports, regulatory filings, contracts, technical manuals and textbooks, who need answers tied to a page or layout block.
  • Developers who already have a model provider and want a retrieval layer not tied to one model, since both modes bring their own LLM.
  • Text-based PDF workloads that keep index files on local disk, since local mode stores documents and indexes in a local directory.
  • Organizations that need private deployment: Enterprise runs inside a customer VPC on AWS, Azure or GCP.

Not a fit:

  • Scanned or image-heavy collections limited to local mode, since the local pipeline runs no OCR.
  • Settings that forbid any document content from reaching a third-party model, unless both the index and chat models are self-hosted.
  • Environments limited to Python below 3.10.

Getting started

Installation. The README gives pip install -U pageindex; the pageindex[claude] extra adds Claude subscription support, and openai and anthropic extras also exist.

Local mode. With the model provider's API key set, a PageIndexClient takes an index model and a chat model, submit_document takes a PDF, and chat answers questions, with streaming available.

Cloud mode. A PageIndex API key from the developer dashboard goes into PAGEINDEX_API_KEY and the index parameter is set to cloud; indexing and storage then run in the cloud while chat still uses the chosen model. The docs also offer four Colab notebooks covering long-PDF indexing, citation highlighting, image retrieval and query cost estimates.

Pricing & license

Local mode is free and open source, and PageIndex bills nothing for it. The rates below apply to PageIndex Cloud only, with usage billing and no monthly minimum:

  • Indexing: $0.01 per page, charged once when the page is indexed.
  • Active pages: $0.001 per page per month, billed at month end, prorated, with the first 1,000 free; deleting a document stops its charges.
  • Retrieval: no fees on active pages; LLM costs go directly to the model provider.
  • New accounts get $10 in free credits with no card required.
  • Earlier Standard, Pro and Max subscriptions keep legacy terms, converting fees to credits at full value with no active-page charges, and move to usage billing on cancellation or change.
  • High volume, custom indexes, higher rate limits, private cloud, SLAs and Zero Data Retention go through sales.

The pricing page's example puts 3,000 pages at $30 one-time indexing and $2 a month after that.

License: LICENSE is the MIT License, copyright Vectify AI. The repository has no CLA and no separate commercial license. The cloud service is also governed by the terms of service.

Risks & limits

Local versus outbound. The Flash blog says Flash runs entirely on the local machine and documents never leave it; the same post says Flash uses the customer's own model API key, and the README says the index model summarizes while the chat model searches the tree.

Scanned files. The Flash blog says the local pipeline reads the PDF directly and runs no OCR, so it expects text-based PDFs, and sends scanned or image-heavy files to the cloud. Issue 106, opened 2026-02-09 23:19 (UTC+8), reports that the tree generation stage is very slow on scanned PDFs when self-hosted, and is still open as of 2026-10-26.

Terms versus current billing. The pricing page, updated 2026-09-18, describes per-page usage billing with no monthly minimum. The terms of service, updated 2024-05-17, describe subscription fees and overage fees per Datapoint, with a GPT-3 completion as the Datapoint example. The free trial clause says data uploaded during a trial becomes inaccessible when the trial ends without a purchase, and the supplier may delete it.

Numbers. The README describes the financial QA benchmark as 34 configurations of 19 LLMs, while the benchmark repository says 35 configurations of 20 LLMs. The README quick-start examples name gpt-6-luna and gpt-6-sol, while the Flash blog and the README benchmark section name gpt-5.6-luna and gpt-5.6-sol. The changelog footer reads last updated 2026-09-18, while its newest entry is dated 2026-09-21.

Issue backlog. As of 2026-10-26 the repository has 51 open issues, and 151 open issues and pull requests combined.

Zero Data Retention. Zero Data Retention on the cloud sits outside self-serve billing and goes through sales.

Our verdict

What sets PageIndex apart: a tree index taken from the document layout, with no vector database and no chunking; answers that cite a page, or a layout block in the cloud; and an MIT-licensed open-source SDK whose local mode is free.

Its limits: both indexing and retrieval call the LLM the user brings; local mode runs no OCR and handles only text-based PDFs; the pricing page describes per-page billing while the terms of service, updated 2024-05-17, still use subscription wording; and the SDK is still at 0.4.x.

Sources

Official

  1. PageIndex - Get verifiable answers from complex documentspageindex.ai · Accessed Oct 26, 2026
  2. PageIndex Developer - The document retrieval layer for AIpageindex.ai · Accessed Oct 26, 2026
  3. About | PageIndexpageindex.ai · Accessed Oct 26, 2026
  4. PageIndex Terms of Service, Privacy Policy and DPApageindex.ai · Accessed Oct 26, 2026
  5. PageIndex Blogpageindex.ai · Accessed Oct 26, 2026
  6. Introducing PageIndex Flash: Fast Local Tree Indexing for PDFspageindex.ai · Accessed Oct 26, 2026
  7. What is PageIndex? - PageIndex Docsdocs.pageindex.ai · Accessed Oct 26, 2026
  8. Pricing for PageIndex Cloud - PageIndex Docsdocs.pageindex.ai · Accessed Oct 26, 2026

Updates & reading