PageIndex Search API
Save your tokens. Send your LLM only the pages that matter.
Upload a PDF, ask up to 10 questions at once, and get the pages that answer each one, every page read, in about a second.
$5.00 of free credit to start, about 10,000 pages.
Two API calls
Upload a document once, then search it as often as you like.
import requests
API = "https://api.search.pageindex.ai/v1"
headers = {"Authorization": "Bearer pis_live_..."}
with open("3M_2018_10K.pdf", "rb") as f:
doc = requests.post(f"{API}/documents", headers=headers, files={"file": f}).json()
result = requests.post(f"{API}/documents/{doc['id']}/search", headers=headers, json={
"questions": ["What was capital expenditure in 2018?", "Who is the CEO?"],
}).json()
for answer in result["results"]:
print(answer["question"], [(p["page"], p["relevance"]) for p in answer["pages"]])Search that reads the whole document
Each page goes to the model as text, or as an image when its meaning is in a chart, a diagram or a form.
About a second
Every page is read at the same time, so a 160-page filing takes about as long as one page. Fast enough for an agent's every step.
Every page, nothing skipped
No embeddings decide what gets looked at. Each page is judged on its own, tables and charts included.
No index to build
No chunking, no vector database, nothing to rebuild when your questions change. Ask anything, any time.
More of the right pages, in a second
Compared with embedding search on the same pages of real 10-K filings.
Evidence pages found
Share of the pages that answer the question, among the first 10 returned
Time to search a new document
A 150-page 10-K and one question, including upload or embedding
80 questions from PageIndex Benchmarks financial-qa, each searched over every page of its 10-K filings (39 filings), with the same page text for every method. Times are approximate, measured from Google Cloud us-central1.
For a search with one question; each extra question in the same search adds $0.20. Uploading and storage are free.
See pricing