Benchmark · Directional
Search-results parsing
Webfold's parser matched OpenSERP's on all 19 saved pages. It was faster on every page. The median gap is 2.6 times. The pages are public regression fixtures. They are not an unseen holdout.
Webfold's results parser against OpenSERP's, on 19 saved pages from six search engines.
What question does this benchmark ask?
Does Webfold read saved search-results pages the way OpenSERP's parser does? How long does it take?
How was it run?
- 19 saved upstream HTML pages from Baidu, Bing, DuckDuckGo, Ecosia, Google and Yandex: results, empty results, captcha and soft-block pages.
- Both parsers called in process, warm, with no subprocess start-up: Webfold's normaliser and OpenSERP's ParseHTML at commit 29c7b0f. Three runs per page, minimum taken.
What are the results?
Total parse time over the 19 pages, millisecondsShorter is better
Webfold
25.7 ms
OpenSERP
93.1 ms
Parsing speed, megabytes per secondLonger is better
Webfold
49.3
OpenSERP
13.6
| Measurement | Webfold | OpenSERP |
|---|---|---|
| Pages whose fields match the oracle | 19 of 19 | oracle |
| Total parse time, 19 pages | 25.65 ms | 93.07 ms |
| Median per-page speed-up | 2.6 times | 1.0 |
| Pages where Webfold is faster | 19 of 19 | 0 |
Webfold matched the oracle's classification and fields on all 19 pages and was faster on every one: 2.6 times at the median, 3.6 times on the total, and up to 5.0 times on a full Baidu results page.
What does this not show?
Know the limits
- These are public upstream regression fixtures, not an unseen holdout, and saved pages say nothing about live relevance or freshness.
- The smallest gaps (1.01 to 1.07 times, on captcha pages) are at timer-noise scale.
- An earlier write-up said 4.7 times. It does not reproduce from the retained files and has been withdrawn; 2.6 times is recomputed from the raw per-page numbers.
- A broad captcha marker can flag ordinary content as a captcha page. The driver mirrors the public parser API; it does not exercise OpenSERP's HTTP or Chromium paths.
- The harness no longer builds from a clean checkout because of a dependency that has moved, so the committed results are not re-run here.
Which other benchmarks are there?
- HTML to Markdown conversion Measured
- Document conversion Measured
- Main-content extraction quality Measured
- Web archive extraction Measured
- Footprint of a scrape or crawl Directional
- PII detection against Presidio Measured
- Localisation extraction Directional
- PDF text extraction with OCR Directional
- PDF extraction against other extractors Directional
- LaTeX conversion with BibTeX Directional
- A whole site to Markdown Directional
- Agent-browser compatibility (Lexbench) Measured
- Headless engine over CDP Directional
- Folder search Directional
Check it on your own pages.
Run the command-line tool on your content.