PageIndex Review (2026): Pricing, Features & Honest Verdict

PageIndex is a document retrieval engine and AI reading tool that answers questions across long professional documents with line-level citations. It is the most credible open-source RAG project we have reviewed this quarter, but the API that made it famous has no published price. Best for: analysts and RAG engineers working with filings and contracts. Price: From $0 (free plan: yes). Rating: 7.5/10.

What is PageIndex?

PageIndex is two products sharing one engine. The first is a retrieval system for developers that drops the entire vector database layer out of RAG. The second is a chat app where you upload a 300-page annual report and ask it questions, with every answer pointing back at the exact line it came from.

The engine is built by Vectify AI and has been public on GitHub since April 2025. We checked the repository through the GitHub API rather than trusting the marketing: 35,364 stars, 3,115 forks, 164 open issues, and an MIT licence. The last commit landed the same day the product hit number one on Product Hunt with 201 upvotes. That MIT licence matters more than it sounds. We have reviewed several products this month that market themselves as open source and ship a noncommercial licence underneath. PageIndex is the real thing.

The pitch is that similarity is not relevance. Standard RAG chops a document into chunks, embeds them, and retrieves the top few by vector distance. PageIndex instead builds a tree index that mirrors the document’s own table of contents, then has the model reason its way down that tree the way a human reader flips to the right section. No embeddings, no chunking, no top-K threshold to tune.

What Are PageIndex’s Key Features?

Vectorless retrieval with no infrastructure to run

The index is a JSON tree. That is the whole storage format. You do not stand up a vector database, you do not maintain an embedding pipeline, and you do not resync anything when a document changes. For a small team this removes a real operational burden, because the vector DB is usually the piece that breaks at 3am. It also removes a cost line that runs into hundreds of dollars a month once your corpus gets large.

Line-level citations you can actually click

Every answer carries explicit page and section references, and in the chat product you click a citation to jump to the highlighted source line. This is the feature that justifies the product for anyone working on filings, contracts or compliance documents. You are not trusting the model, you are checking it in about two seconds. Most RAG tools give you a document name and a confidence score, which is not verification.

Cross-document reasoning and non-text content

PageIndex reads tables, figures, charts and images inside a document rather than skipping them, and it will compare across multiple files in one query. If you have ever tried to get a chunked RAG system to answer “how did this line item move between the 2024 and 2025 filings”, you know that question usually fails because the two numbers sit in different chunks in different documents. Structure-aware retrieval handles it better.

MCP access on every plan, including free

PageIndex exposes itself over MCP, so you can wire it into Claude, an IDE, or your own agent and have the agent search your document library as a tool. The free plan gets basic MCP access and paid plans get full access. Shipping MCP on the free tier is a small decision that says something about how confident they are in the retrieval quality.

How Much Does PageIndex Cost?

PageIndex starts at $0 and the only self-serve paid plan is listed as “from $20 per month”. The free plan gives you 1,000 pages, 100 chat messages, basic MCP access and community support. Pro unlocks full chat, full MCP, priority support and the ability to buy more pages whenever you want, plus a batch of complimentary pages for new subscribers. Team is a custom quote and adds the most accurate model, advanced file search, granular file access controls, multiple users and a private deployment option.

Here is our problem with that board, and it is the main reason this review is not scoring higher. “From $20” is not a price. The page never says what moves the number, and “purchase more pages anytime” appears with no published per-page rate anywhere on the site. You can work out that a page is the billing unit, which is more than several products we reviewed this month managed, but you cannot work out what a page costs once you exhaust the allowance.

The bigger gap is that the developer product has no pricing at all. The engineering page is seven well-argued feature sections, a benchmark chart and a comparison table against vector databases, and then it ends. No tiers, no rate card, just a dashboard link and a documentation link. The API is the reason this project has 35,000 stars, and it is the part you cannot price without booking a call.

For comparison, Ragie publishes a full rate card for managed RAG: $100 per month and $500 per month tiers, extraction at $0.02 or $0.05 per page depending on mode, and storage at $0.002 per page per month. You can model a year of Ragie on a napkin. You cannot do that with PageIndex unless you only care about the chat app.

The free plan is also thinner than it looks. One 300-page 10-K burns nearly a third of your monthly page budget on upload, and 100 chat messages disappears in an afternoon of real analysis. Treat it as an evaluation tier, not a working one.

Who is PageIndex Best For?

Use PageIndex if you work with long, structured, high-stakes documents and you need to show your working. Financial analysts reading filings, lawyers reading contracts, researchers reading papers, and engineers building retrieval over any of those. The citation behaviour is the product. If your users will be challenged on where an answer came from, this is built for exactly that.

Use PageIndex if you are a developer who wants to delete a vector database from your stack. The MIT licence means you can run the engine yourself with no commercial restriction, which makes the missing API price much less of a problem than it first appears. Self-hosting is a real option here, not a theoretical one.

Skip PageIndex if your retrieval need is generic and exploratory. Semantic recommendation, ideation tools, short-document search and broad knowledge lookup are all cases where ordinary vector search works fine and costs less effort. PageIndex says this itself on its comparison page, which is refreshingly honest positioning.

Skip it if you need a fixed, forecastable bill from a managed API this quarter. Until the developer tier is priced in public, budgeting means a sales call.

Best PageIndex Alternatives

Ragie is the direct managed-RAG comparison and the one to look at if published pricing is your deciding factor. Plans at $100 and $500 per month, page-based extraction at $0.02 to $0.05, storage at $0.002 per page per month. You give up the vectorless approach and the MIT-licensed core, and you get a bill you can predict.

LlamaCloud from the LlamaIndex team covers parsing, indexing and retrieval with a free tier to start, and has the deepest ecosystem of connectors and integrations of anything in this space. It is the safe institutional choice. It is also a much bigger surface area to learn, and it keeps you inside the chunk-and-embed model that PageIndex is arguing against.

Building it yourself on pgvector costs nothing but your time and remains the right answer more often than vendors admit. If your corpus is a few thousand documents and your queries are not adversarial, Postgres with an embedding column will serve you for years. Reach for PageIndex when accuracy on structured documents starts failing you, not before.

Final Verdict: Is PageIndex Worth It?

PageIndex is the strongest technical argument we have seen this year for abandoning vector search on domain-specific documents, and the receipts hold up better than most. The 35,364 stars are real, the MIT licence is real, the repository was updated today, and the citation behaviour does the thing the marketing says it does. Their headline benchmark claims 98.7% accuracy on FinanceBench against 30% and 50% for vector RAG, and you should read that as directional rather than independent, because they built the baselines they beat. Even discounted, the architectural case stands on its own.

What holds it back is a pricing page that stops halfway. The chat app is priced, vaguely, and the developer engine that everyone actually came for is not priced at all. If you are an engineer, this barely matters: clone the MIT repo and run it. If you are a buyer who needs a managed API and a budget line, you are booking a demo to learn a number that a competitor like Ragie prints on a public page. Recommended for anyone doing serious document work, with the caveat that you should plan on self-hosting until Vectify publishes an API rate card.