{"slug": "introducing-page-search-api", "title": "Introducing Page Search API", "summary": "VectifyAI released the Page Search API, a retrieval tool that returns whole document pages with relevance scores instead of fixed-size text chunks, so an LLM or agent reads only the pages that answer each question. On 80 questions over 39 10-K filings, Page Search placed 93.2% of answering pages in its top 5, versus 65.5% for Voyage 4 Large, 54.0% for Gemini Embedding 2 and 47.5% for OpenAI text-embedding-3-large, and searched a new 160-page filing in 1.1 seconds against 2.4 to 3.2 seconds to embed the same pages. By default it returns 4.7 pages per question out of a filing's 186, leaving out 97% of the tokens that would otherwise go to the LLM; the API is available through the pageindex Python SDK and over MCP, with benchmark code on GitHub.", "body_md": "Today we are introducing **Page Search API**: give it a document and your questions, and it returns the pages that answer each one, with a relevance score for every page. Your LLM or agent then reads only those pages, not the whole document.\n\nIt is the retrieval step of [PageIndex Agent](https://pageindex.ai/blog/pageindex-agent), offered on its own for teams that already have an LLM, an agent, or a pipeline of their own and want the right pages to feed it.\n\n## \n\nMost retrieval returns chunks: fixed-size slices of text, ranked by how close their embeddings sit to the question. Chunks cut tables, figures and arguments in half, and closeness in meaning is not the same as relevance.\n\nPage Search returns **pages**, the unit a document was written and laid out in. A page keeps its table whole, its chart next to its caption, and its footnote under its claim, and every result is already a citation: page 39 of `report.pdf`.\n\n## \n\nWe compared Page Search with three embedding models, Voyage 4 Large, Gemini Embedding 2 and OpenAI text-embedding-3-large, on 80 questions over 39 annual reports (10-K filings), every method searching the same page text. The measure is the share of the pages that answer that each method returns among its first *k* pages:\n\n|  | Top 5 | Top 10 | Top 20 | \n|---|---|---|---|\n| **Page Search API** | **93.2%** | **97.5%** | **99.3%** | \n| Voyage 4 Large | 65.5% | 83.8% | 90.3% | \n| Gemini Embedding 2 | 54.0% | 70.1% | 86.3% | \n| OpenAI text-embedding-3-large | 47.5% | 68.0% | 79.9% | \n\nIn its top 5, Page Search finds **42% more** of the pages that answer than the best embedding model. For 93.8% of questions, every page that answers is in its first 10 results, against 65.0% for Voyage.\n\nIt is also faster, because there is nothing to build first. Searching a new 160-page filing with one question, upload and processing included, takes **1.1 seconds**, against 2.4 to 3.2 seconds to embed the same pages with the embedding models.\n\nAnd it saves tokens. Left to its default, Page Search lists the pages likely to help: 4.7 pages a question on average, out of a filing's 186. Those few pages hold 95% of the evidence and leave out **97%** of the filing, and so of the tokens you would otherwise send your LLM.\n\nThe questions, documents, results and code are on [GitHub](https://github.com/VectifyAI/page-search-benchmark).\n\n## \n\nInstall the PageIndex SDK, upload a PDF, and ask:\n\n```\npip install -U pageindex\npython\nfrom pageindex import PageIndexClient\n\nclient = PageIndexClient(index=\"cloud\")  # reads PAGEINDEX_API_KEY\ndoc_id = client.submit_document(\"report.pdf\", wait=True)[\"doc_id\"]\n\nresult = client.search([\"What was capital expenditure in 2018?\", \"Who is the CEO?\"], document=doc_id)\nfor answer in result[\"results\"]:\n    print(answer[\"question\"])\n    for page in answer[\"documents\"][0][\"pages\"]:\n        print(\"  page\", page[\"page\"], page[\"relevance\"])\n```\n\nAsk up to 10 questions in one search; ten take about as long as one.\n\n## \n\n- **Pictures too.** The search is multimodal: it reads scans, photos, charts and diagrams, so you need no OCR model or PDF parser.\n- **One document or many.** Search one document, a collection, documents matching a metadata filter such as`{\"ticker\": \"NVDA\", \"year\": 2026}` , or all of them.[Search documents](https://pageindex.ai/docs/search/search-documents) finds which documents answer before you search their pages.\n- **Built for agents.** The SDK gives your agent ready-made search tools, and the API is also available over MCP.\n- **One account with Page Extract.** The same key, credit and documents work with[Page Extract API](https://pageindex.ai/blog/page-extract-api) : upload once, then search a document or extract from it.\n\n## \n\nCreate a key in the [Developer Dashboard](https://developer.pageindex.ai/api-keys) and follow the [Quick start](https://pageindex.ai/docs/search/quick-start), or paste this into Claude Code, Cursor, or any coding agent that can read a URL:\n\n```\nRead https://pageindex.ai/docs/search/quick-start.md and follow it to set up Page Search API in this project.\n```\n\nSee [pricing](https://pageindex.ai/pricing), or read more on the [Page Search API page](https://pageindex.ai/search).", "url": "https://wpnews.pro/news/introducing-page-search-api", "canonical_source": "https://pageindex.ai/blog/page-search-api", "published_at": "2026-11-14 11:00:00+00:00", "updated_at": "2026-11-16 04:20:09.326579+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-search", "large-language-models", "agent-protocols"], "entities": ["VectifyAI", "Page Search API", "PageIndex", "Voyage 4 Large", "Gemini Embedding 2", "OpenAI text-embedding-3-large", "GitHub", "MCP"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/introducing-page-search-api", "markdown": "https://wpnews.pro/news/introducing-page-search-api.md", "text": "https://wpnews.pro/news/introducing-page-search-api.txt", "jsonld": "https://wpnews.pro/news/introducing-page-search-api.jsonld"}}