Skip to content

How a link reads itself

Markdown

Save a link and Copilka reads the page behind it. A moment later the card has a title, a picture, a summary and tags, and, when the page is a recognisable kind of thing, the facts that go with that kind: the price of a product, the date of an event, the seat on a ticket. This page explains where those come from, in what order, and what you can do when the reading is not what you wanted.

  • While you type an address into the Save a link sheet, Copilka reads the page quickly, from the sites it knows and the page’s own markup only, so the title is there before you press Save. The sheet says Reading the page… meanwhile.
  • The moment you save, the full reading starts. Saving never waits for it: the card appears at once and fills in as the reading arrives. Things you save yourself start straight away; anything waiting from earlier is read four at a time in the background.
  • On launch, Copilka reads anything never read, and, depending on Try again on launch, retries pages that said nothing last time.

Two settings in Settings → Reading govern this:

  • Read new saves: Fetch and read a link the moment it lands. Off, nothing is fetched until you ask. On by default.
  • Try again on launch: Never, Pages that said nothing (the default: links whose reading came back with no title and no picture, more than ten minutes ago) or Oldest readings too.

Notes and videos are never read. Pictures and files are; see Pictures.

A page already in the library is not read again. The new copy adopts the existing one’s reading, price history and change history at once, offline. If you later use Read this page again on either copy, the fresh reading reaches every copy of that page.

Each step fills in what the previous ones left empty; nothing later overwrites what an earlier step found. Reading stops as soon as the result is enough for the kind of thing it is: an event needs its date, a product needs a price, availability or a barcode. The whole reading has a deadline of about a minute.

  1. Sites Copilka knows directly. For many sites Copilka asks the site’s own service rather than reading the page: maps, YouTube and Vimeo, GitHub, GitLab and Codeberg, arXiv, package registries, the App Store, podcasts, Stack Exchange, Hugging Face, Spotify and the other music services, Steam, and the social networks. See Kinds of things for the list. For book, event and film sites, Copilka knows the shape of the thing before it has read a word, so the tile is the right one from the start.
  2. The page itself. The page is fetched with a ten-second timeout. Copilka reads its structured data first (JSON-LD, microdata, OpenGraph), then the shop-specific hints it knows, then the title, description and icon. It also keeps the readable text of the page, up to about twelve thousand characters, with menus, footers and forms stripped out, and stores an excerpt of up to six thousand characters with the reading. A page that answers 404 or 410 is marked as gone.
  3. A model. If the page’s own data was not enough and there is at least a paragraph of readable text, the text is sent to a model together with the address, the title, the hints already found and any Where to look note you have left. The model decides what kind of thing the page is and pulls out the facts for that kind. This step needs the subscription: see Plans and invites.
  4. A browser. If the page looks like an app shell or a bot wall and gave nothing away, Copilka loads it in a real browser and reads it again. It tries a browser on this Mac first, then one carrying your own sessions if you have turned Read as you on, and only then a hosted browser run by Copilka, for that one address. This step also needs the subscription.
  5. Tags. Up to five tags come from the page’s own data. If there are fewer, and a model is allowed, it is asked for up to three more, each with an emoji, and it is shown your existing tags so that it reuses your vocabulary rather than inventing one. A new tag is filed under an existing parent when that makes sense.

Readings are cached for a week, so re-saving a page you removed does not fetch it again. A page that failed is tried again after ten minutes.

  • The site name, icon and picture are filled in only where they are empty.
  • Tags are added, never removed.
  • The title is replaced only if you have not edited it yourself, and only by a reading at least as well-founded as the one before.
  • A blank reading never erases a good one. A page that did not answer this time keeps what it said last time.
  • A sign-in wall lends nothing. If the page came back as a login page, an access-denied page or a “just a moment” check, the reading is kept for the record but no title, picture or tags are taken from it, and the item offers to sign in instead. See Read as you.
  • Prices and other watched facts are recorded for the history: see Prices and changes.

While a page is being read the card’s name is a small animated grid of squares, the summary in the details panel is three grey lines, and a link tile’s picture area shimmers gently. When the reading lands, a new title and summary type themselves in over a moment. Appearance → Show new text being written turns that off: Titles and summaries arrive like a reply after a page is read.

Reading again, and telling the reader where to look

Section titled “Reading again, and telling the reader where to look”

In the details panel, Read this page again opens a small sheet:

  • Read again reads the page afresh, skipping the cache. The toast says Reading … again.
  • Where to look is a note to the reader, kept on this item and sent with every later read. The placeholder shows the idea: The seat is the box at the bottom right, not the row number. For a link, the note is only consulted when the model is, so it helps on pages whose own markup fell short.

The same sheet on a picture or a file also has a What it is menu; see Pictures.

Reading a page again can replace a title you typed, since it starts the reading over. Edit the title afterwards if you want yours back. Individual facts you disagree with can be overridden as properties, and your value stands even when the page changes.

The Developer → Enrichment Inspector window (⌥⌘E) takes any address and runs the reader on it in front of you, fast or in full, showing what it resolved through, what it did, and the raw result. Nothing you paste there is saved.