Why Sluicer

Most of what a scraper is asked for is already written into the page it scrapes. A product page carries its name, price, currency and SKU in JSON-LD for search engines; an article carries its headline, author and dates in OpenGraph and meta tags. Sluicer reads that, rather than guessing it from the visible text or asking a model to.

This page says where that puts it among the tools people reach for, and where another tool is the better choice. It is a map, not a race.

Where it sits

	Sluicer	extruct	trafilatura	CSS-selector scraper	LLM scraper
Reads declared data (JSON-LD, microdata, RDFa, OpenGraph...)	yes, merged	yes, per vocabulary	partly	if you write it	if the model notices
Reads authors and dates from the visible prose	no	no	yes	if you write it	yes
One record per thing, across vocabularies	yes	no	no	no	varies
Says where every value came from	yes, reader and key	per vocabulary	no	no	no
Same page, same answer	yes	yes	yes	yes	no
Notices when a site's layout changes	yes, exit 3	no	no	no, returns nulls	no
Says what moved after a redesign	yes, heal	no	no	no	no
Fetches JavaScript-rendered pages	yes, a browser when measured necessary	no	no	depends	depends
Cost per page	CPU	CPU	CPU	CPU	tokens

"Partly", "varies" and "depends" are honest: trafilatura reads some metadata tags on its way to the text, an LLM scraper's behaviour is its prompt's, and a hand-written scraper does whatever its author wrote.

What Sluicer adds

One record per thing, with its provenance. A page that describes the same product in JSON-LD, microdata and OpenGraph gives one record, and every field still names the vocabulary that declared it. Folding happens across vocabularies, never inside one: two products on a listing stay two products, because folding them would splice one's name onto the other's price.

A summary that answers the usual questions, and says how. Title, description, author, dates, image, language, site name, publisher, type, price, currency, availability, brand and SKU, one value each, chosen by fixed rules -- the article's headline before the site's name, og:title before <title>, the price from inside offers -- and each naming its reader and key, so it can be checked against the records.

The tags nobody owns. Across the 359 commercial pages of a public annotated corpus, article:published_time is on 33% and <meta name="author"> on 29%. The first is OpenGraph's article: namespace; the second belongs to no vocabulary, and Sluicer reports it as "source": "html", because that is what it is.

Extractors that fail loudly. Learn an extractor from two or three pages of a template, replay it on any page of that template, and a page that drifted -- the listing moved, a field emptied, a price slot now says "Add to basket" -- fails with exit code 3 and the check that broke, instead of returning nulls for weeks. After a redesign, heal says which field moved where, how many of its old values were found in the new place, and keeps the column names your code reads. See extractors.

A fetch that announces itself. Plain HTTP first, a browser only when a measurement says the cheap rung got a refusal, a challenge or an empty shell, and every climb reported with its reason and its cost in seconds. Every request says Sluicer/<version>, borrows no browser's referer or fingerprint, and obeys robots.txt -- and nothing is fetched when robots.txt cannot be read, as RFC 9309 says. The stealth rung exists and never runs unless asked.

A tool an agent can trust. Six MCP tools, each answer with ok and an output schema, an error that can never be mistaken for the page, and a server that keeps every request -- redirects, images, frames, websockets -- off private addresses unless told otherwise.

When to use something else

• You need the author or date of pages that do not declare them. trafilatura reads them from the visible text, and on the WCXB corpus it finds far more of them. It also dates 216 pages that have no date; Sluicer answers less often and is wrong less often. See the scoreboard.
• You need the article's text. sluicer markdown uses trafilatura for exactly that; for anything beyond it, use trafilatura directly.
• The data is not declared and the pages are not a repeated template. Induction reads listings and feeds; it does not read an arbitrary page's prose into fields. That is a model's job, and its answers should be checked.
• You are crawling a site. Sluicer fetches the page you ask for; it does not follow links, pace itself by Crawl-delay, or keep a frontier. Put it behind a crawler.
• You only need one vocabulary, raw. extruct returns each vocabulary as the page wrote it, which is what you want when the merge is not.