Spelling Creator

Search images ​

Press Search images on any section to open a dialog that searches for free images and inserts the one you pick as an image block. It has two sources, switched with the toggle at the top: Pixabay and Wikimedia Commons. Both work the same way from there: search, click a result, and it is inserted with its caption pre-filled with the attribution the source asks for, editable like any other caption.

Each source sits behind the same small interface in ImageSearchDialog.jsx (search, resolve, caption), so the dialog's flow doesn't care which one is selected.

Pixabay ​

Pixabay goes through the companion Worker (apps/api) rather than being called directly, which:

  • keeps the Pixabay API key server-side (it is never shipped to the browser),
  • lets the Worker enforce Pixabay's 100 requests/minute limit centrally, and
  • works around Pixabay's image CDN sending no CORS headers: the browser can't read an image's bytes itself, so the Worker downloads the chosen image and returns it as a data URL that drops straight into the DOCX/PDF export.

The flow:

  1. Type a search term; a Cloudflare Turnstile token (the same widget as the AI dialogs) is sent with the request.
  2. The Worker verifies the token, calls the Pixabay API with mode: "imageSearch", and returns normalised hits (preview/webformat URLs, size, tags). It edge-caches the Pixabay response for 24 hours, which both satisfies Pixabay's caching requirement and keeps request counts well under the limit.
  3. Click a result; the app calls the Worker again with mode: "imageFetch", which downloads that image and returns it as a data URL.
  4. The image is inserted with the caption Image by {photographer} from Pixabay.

Each Worker call consumes its single-use Turnstile token, so the widget is reset to mint a fresh one between searching and inserting. This source needs the same VITE_API_URL / VITE_TURNSTILE_SITE_KEY as the AI features, plus a PIXABAY_API_KEY secret on the Worker (wrangler secret put PIXABAY_API_KEY).

Wikimedia Commons ​

Commons is called straight from the browser (@spelling-creator/core/browser/commonsImages). Its API answers anonymous cross-origin requests and its image CDN sends CORS headers, so there is no key, no Worker and no Turnstile challenge. The MCP server's search_images searches Commons too, with the same shared plumbing (@spelling-creator/core/wikimedia).

Every Commons file is licensed on its own, so each result carries its author and licence, and the caption is built from them: Image (by {author}, {licence}) via Wikimedia Commons.

Wikidata's pictures come first ​

When the search names one particular thing ("lion", "Paris", "Great Pyramid of Giza"), the results open with the pictures Wikidata lists for it, each with a badge saying what it is: Main picture, From above, At night, Map, Where it lives, Flag, and so on. A line above the grid names the item they are from, with its description, so a search that landed on the wrong "Mercury" says so before anyone picks from it.

Those pictures were chosen by the people describing that thing on Wikidata, so they are usually a better first offer than whatever a full-text search of Commons ranks highest. They are ordinary Commons files, so everything after the search (the download, the caption) is the same code. A picture that is also in the search's own results isn't shown twice.

How the item is found (packages/core/src/wikidataMedia.js):

  1. Only an exact name counts. wbsearchentities matches prefixes, so the lookup keeps only items whose label or alias is the search. "Volcano erupting" names no one thing and gets no Wikidata pictures, rather than the wrong thing's.
  2. The best known match with pictures wins. Wikidata's search order is a poor guide: for "lion" it puts a family name first, and for "Mercury" a car brand ahead of the planet. So all the exact matches go into one SPARQL query that returns their pictures and how many Wikipedias have a page on each (their sitelinks). The lion has 274 and the family name 2.
  3. Pictures are read by property, in this order: main picture (P18), aerial view, night view, panorama, interior, montage, range map, locator map, location map, flag, coat of arms, logo, chemical structure. Deprecated statements are skipped, and so are ones with an end date (Japan's flag from 1870 to 1999). At most two of any kind (Japan has five locator maps) and eight in all.

Some picture properties are left out on purpose, such as an image of a grave or a seal: real pictures of the thing, but not ones to put in front of a speller unasked.

Wikidata is an extra here, not the search. If it is slow or down, the dialog still shows Commons' own results and says nothing about it. The lookup is three small requests (a name search, one query, and a Commons lookup of the files), made alongside the Commons search rather than before it, with a budget of four seconds for all three (wikidataPickPages in packages/core/src/wikimedia.js, which both apps use). A query service under load costs the picks, never the search.

A file Wikidata names may have been renamed on Commons since, with a redirect left behind, or be spelt with different capitals. The lookup asks Commons to follow redirects and maps each title through the renames Commons reports, so the picture still shows under its current name.

The shared Wikidata plumbing (searching names, running queries, reading Commons file names out of query results) is packages/core/src/wikidata.js, which the fact check uses too.

Copyright © 2026 Spelling Creator.