Introducing the New PageIndex Agent

Vectorless, OCR-free, Multimodal RAG
Published on

Today we are introducing the new PageIndex Agent: vectorless, OCR-free, multimodal RAG for complex documents, with line-level citations for every answer.

PageIndex Agent is what we used to call the PageIndex SDK and Cloud, rebuilt and renamed. It is now one of three PageIndex products, alongside Page Search API and Page Extract API, and it is the one to use when you want answers: you bring your documents and your own LLM, and PageIndex finds the evidence and cites it.

How it finds an answer

Vector RAG splits documents into chunks and retrieves the ones whose embeddings sit closest to the question. Closeness in meaning is not the same as relevance, and in long financial filings, contracts and manuals the passage that answers often shares few words with the question.

PageIndex Agent searches the way an expert reader does, in four steps:

Annual report120 pagesBusiness overviewp. 1–24Financial resultsp. 25–78Risk factorsp. 79–120p. 25p. 26p. 27p. 28p. 29p. 30p. 31p. 32p. 33p. 34p. 35p. 36p. 37p. 38p. 39your documentsFinancial results, p. 25–78p. 341.Find relevant documents2.Find relevant sections3.Find relevant pages4.Find it on the page
Annual report120 pagesBusiness overviewp. 1–24Financial resultsp. 25–78Risk factorsp. 79–120p. 25p. 26p. 27p. 28p. 29p. 30p. 31p. 32p. 33p. 34p. 35p. 36p. 37p. 38p. 39your documentsFinancial results, p. 25–78p. 341.Find relevant documents2.Find relevant sections3.Find relevant pages4.Find it on the page
  1. Find the relevant documents. Keep documents in collections tagged with your own metadata, or in folders, and each question goes only to the documents that match.
  2. Find the relevant sections. PageIndex builds a tree of each document's sections, like a table of contents, and the agent navigates it as a person would: straight to the sections that matter, each traced to its pages.
  3. Find the relevant pages. Page Search narrows each section to the few pages that answer, reading text and images alike.
  4. Find it on the page. The agent reads those pages visually and outlines the evidence it used, down to the line.

It never reads the whole document, so it spends far fewer tokens, and every step can be checked.

What's new

Page Search inside

Step 3 is new. Once the tree has picked a section, Page Search scores every page in it against the question and keeps only the pages likely to answer. The agent then reads a handful of pages instead of a whole section.

Page Search is the same engine we sell as its own API, and on our benchmark of questions about 10-K filings it finds 99.3% of the pages that answer in its top 20 results, against 90.3% for the best embedding model we tested (Voyage 4 Large). It searches a new 160-page document in 1.1 seconds, with no indexing or preprocessing, and the pages it returns are about 97% fewer tokens than the whole document.

Multimodal, OCR-free

The new PageIndex Agent reads pages as they are, not as text pulled out of them. Charts, tables, scanned pages and engineering drawings are read visually, so a value that lives in a chart or a drawing is as findable as one in running text, and nothing is lost to an OCR step first.

Citations follow the same idea. Each one carries the page and the block it came from, with its position on the page, so you can outline the evidence on the page rather than only name a page number:

citations[0]
# {'document': 'report.pdf', 'page': 12, 'block_id': 'p12_text_3',
#  'bbox': [78, 24, 293, 44], 'block_type': 'text', 'text': '…'}

Collections with metadata, or folders

Documents can now be organized in two ways, whichever fits your project:

  • Collections with metadata: group documents and tag each with your own keys, such as company or year, so a question goes only to the documents that match.
  • File system with folders: keep documents in folders and subfolders, and ask within a folder.

Either way, step 1 of the search is the organization you already have. See Collections with metadata and File system with folders.

One of three products

PageIndex now comes as three products built on the same document understanding:

You wantUse
Answers to questions about your documents, with citationsPageIndex Agent
The relevant pages, to feed your own LLM or agentPage Search API
Fields and tables as JSON, to store or processPage Extract API

PageIndex Agent also powers our own Chat App and Page Extract.

Bring your own LLM

PageIndex indexes and stores your documents; your own model searches the tree and writes the answer. That keeps you free to pick the balance of accuracy and cost that suits you, and to switch models as they improve, without re-indexing anything:

from pageindex import PageIndexClient

client = PageIndexClient(
    index="cloud",            # index and store in PageIndex Cloud
    chat="gpt-6.1-terra",     # your own model answers
)

doc_id = client.submit_document("annual-report.pdf", wait=True)["doc_id"]

for chunk in client.chat("What drove revenue growth?", doc_id=doc_id, stream=True):
    print(chunk, end="", flush=True)

chat() runs PageIndex's own document agent with your model. To use your own agent instead, take PageIndex's retrieval tools and system prompt into the framework you already use, in Python or JavaScript, or connect over MCP. See Use with your LLM and Use with your agent.

Get started

Install the SDK, create an API key, and ask your first question in a few minutes:

pip install -U pageindex

Or paste this into Claude Code, Cursor, or any coding agent that can read a URL:

Read https://pageindex.ai/docs/SKILL.md and follow it to set up PageIndex in this project.

Read the Quick start, see the PageIndex Agent page, or try it on real documents in our examples.