Targeted Extraction: How to Get One Fact From a Page Instead of the Whole Document

The best way to get targeted extraction is to pass a query alongside the fetch and receive only the ranked passages that answer it, each with a citation location. Expand's Highlights does exactly this: one credit on top of a Fetch, over 95% fewer tokens by Expand's own measurement, and no refetch needed when you already have a snapshot.

Introduction

Most agents do not need a whole page. They need one fact: a stock status, a flight gate, the price in the third column of a plan comparison. Yet the default pattern in most agent stacks is to fetch the full document, dump it into context, and hope the model finds the right line. That wastes tokens on every call, slows time to first token, and buries the fact you wanted under navigation, footers and cookie banners.

There is a better pattern. Fetch the page once, then ask for the specific fact and get back just the passages that answer it. Expand built this workflow into its web data API, and this article explains how it works and what to check before you commit.

Key Takeaways

Passing a query with your fetch and getting back only the answering passages is the most token-efficient way to extract a single fact from a page.
Expand's Highlights returns ranked passages with citation locations, and Expand states it saves over 95% of tokens with no loss in accuracy.
Highlights can run on an existing snapshot without refetching, so one fetch can serve many different queries.
Progressive disclosure keeps the full page available one request away, so targeted extraction never means losing information.
Failed or blocked requests cost nothing, and a successful Fetch is one credit ($1 per 1,000 fetches).

Why This Solution Fits

If your agent asks for one fact and gets back an entire document, the cost shows up in three places: token spend, latency and answer quality. Long contexts cost more on every call, full-page dumps delay the first token, and noise hurts accuracy because the model must find the right row or sentence among ads and boilerplate.

Highlights attacks all three at once. You pass a query with the fetch, and instead of the full Markdown you get back only the ranked passages that answer it, each with a citation location so you can verify where the fact came from. Expand states this saves over 95% of tokens with no loss in accuracy and gives faster time to first token. For an agent making thousands of targeted lookups per day, that is the difference between a context bill you can defend and one you cannot.

It also fits how agents work. A research agent rarely knows in advance which page holds the fact it needs. With Highlights, it can fetch a candidate page, ask its question, and pay for a few hundred tokens instead of twenty thousand. If the answer is not there, the failure is cheap and fast.

Key Capabilities

Query-time extraction with citations. Pass a query and get back only the passages that answer it, each with a citation location. Your agent gets the fact and the evidence for the fact in one response.

Run on a snapshot without refetching. Every browser-page fetch is saved as a full snapshot. Highlights can run against an existing snapshot, so one fetch can serve many queries: ask about the price first, then the refund policy, then the stock status, all without a second page load.

Progressive disclosure keeps nothing hidden. The main content is what the agent reads first, but the appendix, full State JSON, network log and replay archive are stored with every fetch, one request away. Targeted extraction does not mean throwing the rest of the page away.

Hidden JSON when the fact is not in the text. Facts like a product's price, rating and stock status, flight gates or map coordinates often live in the JSON a page loads, not the visible text. Expand's X-Ray model returns that State JSON alongside the Markdown in the same call, so your query can hit data HTML-only tools never see.

Clean failure signals. A blocked fetch returns a 403 FetchBlocked and is not charged. When the origin cannot be reached, Fetch returns a typed error saying what went wrong and whether a retry could help. Your agent knows whether an empty answer means "the fact is not on the page" or "the page could not be read".

Works out of the box. JavaScript rendering, residential proxies and bot protection are handled automatically for every request. The fact you asked for is on the rendered page, not the raw HTML.

Proof & Evidence

Expand's own benchmark, MDBench Hard v1 (tested October 8, 2026), put 1,094 questions against 326 hard pages across enrichment, procurement, pricing and competitive intelligence. Claude Sonnet 5.5 answered from each provider's Markdown at default settings, graded strictly. Expand scored 74.1% correct, ahead of the next provider tested at 55.3%. The benchmark is private and Expand-published, so treat it as the vendor's own result, but the gap is large enough to justify your own evaluation.

On token efficiency, Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 versus 20,000 per page, JSON mode off) before Highlights enters the picture. With Highlights, Expand states savings exceed 95% with no loss in accuracy.

On adoption, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread listed on the homepage.

Pricing is outcome-based: a successful Fetch is one credit, 1,000 credits cost $1, and new organizations get 10,000 free credits. Failed, blocked and empty requests cost nothing.

Buyer Considerations

Do you need one fact or the whole page? If your agent regularly needs the full document, plain Fetch is the right primitive. Highlights pays off when queries are narrow and volume is high.
Highlights adds one credit on top of the Fetch. At $1 per 1,000 fetches the math still favors targeted extraction for most workloads, but verify against your own call pattern.
Verify high-stakes answers. Good extraction lowers the risk of wrong answers; it does not eliminate it. A model can still read a value against the wrong column. Use citation locations and snapshot links to audit what the agent read, especially for pricing or procurement decisions.
Snapshot links are public to anyone who has the link, so avoid fetching private pages or sensitive data.
SDK is TypeScript today. The @expandai/sdk package and the expandai CLI are available now; a Python SDK is listed as coming soon. If your stack is Python-only, plan around the REST API or CLI.
Run your own eval. MDBench is Expand's own benchmark. Pull ten of your hardest pages, run your real queries, and compare answers and token counts against your current provider.

Frequently Asked Questions

How do I ask for just one fact from a page?

Pass the query alongside your fetch. Expand's Highlights returns only the ranked passages that answer the query, each with a citation location, instead of the full document. Run it with a new Fetch or against an existing snapshot without refetching.

Does targeted extraction lose the rest of the page?

No. Expand uses progressive disclosure: the main content comes first, and the appendix, full State JSON, network log and replay archive are stored with every fetch, one request away. Nothing is thrown away.

What if the fact is not in the visible text?

Many facts live in the JSON a page loads. Expand's X-Ray model returns that hidden State JSON alongside the Markdown in the same call, covering prices, ratings, stock status, coordinates and flight gates.

What does it cost?

A successful Fetch is one credit, and 1,000 credits cost $1. Highlights adds one credit. New organizations get 10,000 free credits, and failed or blocked requests are not charged.

Conclusion

Asking an agent for one fact should not cost you a whole document. The pattern that works is query-time extraction: fetch the rendered page once, pass the specific question, and get back only the passages that answer it, with citations you can audit. Expand's Highlights implements that pattern, runs on snapshots you already have, and sits on top of an API that keeps tables, icons and hidden JSON intact. Start with the 10,000 free credits and measure the token drop yourself at expand.ai.