You.com's Extract API Turns Broken PDF Tables Into Agent-Ready Data
You.com launched the Extract API , which fetches and parses any web page or PDF into markdown or HTML. On ParseBench-2 it averaged 78.6, versus 43.2 for Exa and 38.3 for Parallel: about 80% higher than Exa. The biggest gaps are on tables and charts.

Takeaways
- You.com launched the Extract API , which fetches and parses any web page or PDF into markdown or HTML.
- On ParseBench-2 it averaged 78.6, versus 43.2 for Exa and 38.3 for Parallel: about 80% higher than Exa.
- The biggest gaps are on tables and charts. Extract estimates values for chart bars that have no printed labels.
- Pricing: $0.90 per 1,000 URLs. The default standard mode adds $0.50 per 1,000 PDF pages.
- Tradeoffs: 11-second median latency in standard mode, and LlamaParse still scores higher on the benchmark .
- Free for web pages via the MCP server; new signups get $10 credit, or $100 with a work email.
Web-extraction APIs often pull embedded text from PDFs while losing the document’s structure. Table values drift away from their headers, charts collapse into axis labels, and checkboxes become detached from their fields. Those errors can corrupt answers drawn from filings, research papers, forms, and other structured documents.
You.com has launched the Extract API to preserve those relationships. The API accepts a web page or PDF URL, fetches the source, and returns markdown or HTML. It combines retrieval and parsing in one request, giving agents access to complete documents after search or from URLs supplied directly.
You.com’s main evidence is a company-run ParseBench-2 evaluation. Extract’s standard mode averaged 78.6 across five scored dimensions, about 82% above Exa Contents at 43.2 and 105% above Parallel Extract at 38.3. Those percentages describe relative gains, not percentage-point differences.
One call replaces fetch-then-parse
Document pipelines commonly use one service to fetch a URL and another to parse any PDF it returns. Extract performs both operations through one endpoint and offers two PDF processing levels through the parse_effort parameter. The parameter does not affect web pages.
standard: Runs PDF pages through deeper parsing intended to recover tables, charts, images, forms, and other structured content. This is the default.lite: Returns text embedded in each PDF page without the deeper parsing pass.
Extract serves cached content when available and fetches the source when no cached copy exists. Setting extraction_source=fetch forces a live request, which is useful when a page changes frequently or freshness matters more than response time.
Charts become estimated data points
PDF charts may contain vector drawings with axis labels but no embedded values. Extract analyzes each bar against the axes and estimates its value. In You.com’s example from an International Energy Agency report, a page contained six bar charts with 12 values. Extract returned a table for each chart and received a score of 100; Exa and Parallel returned axis ticks and received 0.
The company’s other published samples test whether the parser preserves relationships that determine a document’s meaning:
- Tables: Extract scored 100 on an eight-column Deutsche Bank liquidity table. Exa shifted values away from their headers, while Parallel produced unreadable characters; both received 0.
- Forms: Extract scored 95 on a completed USCIS form. Exa and Parallel scored about 50 after associating a checkmark with the wrong field, which would cause an agent to identify the wrong parent.
- Handwriting: Extract scored 95 on photographed Czech lecture notes and represented the boxed formula as renderable LaTeX.
You.com has published the tools’ raw outputs on GitHub, allowing developers to inspect the examples and rerun the evaluation.
Accuracy comes with an 11-second median
ParseBench-2 is a document-parsing benchmark created by LlamaIndex. Its severity-aware grader assigns larger penalties to errors that are more likely to break downstream applications. A displaced table header therefore costs more than an extra space. You.com evaluated 3,782 pages across 12 sets using the benchmark’s grader and excluded its layout dimension because Extract returns markdown or HTML without page coordinates.
| Metric | Extract standard | Extract lite | Exa Contents | Parallel Extract |
|---|---|---|---|---|
| Tables | 74.9 | 23.9 | 22.8 | 0.1 |
| Charts | 76.6 | 1.5 | 0.9 | 0 |
| Forms | 75.5 | 65.4 | 62.6 | 61.0 |
| Five-dimension average | 78.6 | 46.5 | 43.2 | 38.3 |
| Median latency | 11.0 s | 2.6 s | 1.3 s | 1.8 s |
Lite’s 46.5 average sits near the competing services, indicating that the deeper PDF pass produces most of standard mode’s accuracy gain. It also raises median latency to 11 seconds per page, compared with 2.6 seconds for lite, 1.3 seconds for Exa, and 1.8 seconds for Parallel. That delay is easier to absorb in asynchronous batch processing than in an interactive agent response.
Page fees change the cost equation
Extract charges $0.90 per 1,000 URLs in lite mode for web pages and PDFs. Standard processing adds $0.50 per 1,000 PDF pages, so document length determines the additional charge.
You.com reports a median processed PDF length of 26 pages. At that length, 1,000 standard-mode PDFs cost $13.90: $0.90 for the URLs plus $13 for 26,000 processed pages. According to You.com’s comparison, Exa and Parallel each charge $1 for 1,000 such PDFs because their billing is based on URL count.
Three limits behind the benchmark claim
- Dedicated parsers score higher: You.com acknowledges that LlamaParse leads Extract on ParseBench-2. LlamaIndex’s paper reports an overall score of 90.1 for LlamaParse Agentic Plus and 85.7 for the strongest tested frontier vision-language model. Benchmark totals should be compared only when the evaluated dimensions and settings match.
- Layout was omitted: ParseBench-2 can score the position of content on a page. Extract does not return page coordinates, so You.com removed that dimension from its evaluation.
- The evaluation is self-reported: You.com ran the tests and selected the showcase examples. LlamaIndex created the benchmark and sells the parser at the top of its leaderboard. The public dataset, grader, and raw outputs make independent replication possible.
LlamaIndex CEO Jerry Liu has argued that aggregate scores can obscure workload-specific errors. A back-office forms pipeline may prioritize field associations and table headers, while a research workflow may assign more weight to equations or charts. Teams should compare the relevant dimension scores and test representative documents before selecting a parser.
From first request to migration
Developers can test web-page extraction through the You.com MCP server without an API key or authentication. Direct API access includes $10 in free credit, increased to $100 for sign-ups using a work email. Zero data retention, which prevents submitted content from being stored, is available at no added cost after You.com enables it for the account.
Standard mode suits filings, earnings reports, and papers whose answers depend on tables, charts, forms, or images, particularly in batch pipelines that can tolerate higher latency. Lite mode fits text-based PDFs such as contracts and interactive workflows that need faster responses.
You.com will continue supporting its older Contents API, while deeper PDF parsing, parse_effort , and future extraction capabilities are assigned to Extract. The open ParseBench-2 dataset provides a starting point for independent tests, though production evaluations should include the documents, error types, latency limits, and page counts expected in the target workload.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

