Files
karl b8400c4fbf
Core CI / core (ubuntu-latest) (push) Failing after 3m31s
Core CI / core (macos-latest) (push) Canceled after 0s
Core CI / core (windows-latest) (push) Canceled after 0s
Core CI / desktop (ubuntu-latest) (push) Successful in 5m39s
Core CI / desktop (macos-latest) (push) Canceled after 0s
Core CI / desktop (windows-latest) (push) Canceled after 0s
feat: add multimodal search and granular performance diagnostics
2026-10-12 19:16:23 +02:00
..

Independently labelled media diagnostic v1

fixture.json and labels.lock.json were frozen before media inference. The fixture contains 20 original public/licensed files and 65 English/German queries. Neutral filenames prevent filename semantics from giving away visual answers. Do not edit v1 labels after seeing rankings; corrections require a new version and another pre-scoring lock.

The assistant inspected every photo and sampled real video frames through view_image. Cat-action/flower-state intervals came from these frames. Spoken-topic intervals came from NASA's publisher-provided SRT, not a model transcript. Music/sound labels came from independent publisher descriptions and original embedded tags; human listening verification is not claimed. One annotator authored these labels, so a blinded second human review remains necessary.

Provenance and redistribution

Every asset has its exact source URL, description page, license, byte count, SHA256, and modification record in the fixture. The assets are a small educational search benchmark, not a stock-media product.

  • Twelve 640px photos are publisher-generated derivatives under the Pexels License; no local pixel modifications were made. Pexels IDs identify the sources; attribution is not required by that license. Creator names are not guessed.
  • clip01.webm: Cat opening door, Astropulse, source, CC BY-SA 4.0. The original is unmodified. Reproduced frames are derivatives and retain that license and attribution.
  • clip02.webm: Water flowing strem, BhotaRohit, source, CC0. Visual inspection reveals a misty waterfall; labels deliberately follow the pixels rather than the filename.
  • clip03.webm and sound02.mp3: MDN contributors, original interactive-example assets, repository CC0 notice. The files are the real flower timelapse and roaring-creature effect.
  • clip04.mp4: NASA Artemis Origins, source, used for educational/informational evaluation under NASA media usage permission. No endorsement is implied; NASA insignia/trademarks are not used as Searchi branding. The footage is unmodified.
  • music01.ogg: Variatio 13 a 2 Clav., Kimiko Ishizaka, source, CC0. The project's primary release description confirms the public-domain release, independently of the Commons pending-review banner. Original audio and Vorbis metadata are retained.
  • music02.ogg: Jazz Violin Solo, VWD, source, author's worldwide public-domain dedication on that page. Original tags identify the artist as Violinist WTD, which is intentionally the strict metadata query rather than a guessed identity.
  • sound01.wav: Dog barking mono, Brandon Morris, uploaded by HaelDB, source, CC0. Original waveform retained.

All source licenses remain attached to their files. Root repository MIT licensing does not replace these third-party media terms. derived/ and inspection/ are ignored; generated derivatives carry their source licenses when redistributed.

This collection is deliberately tiny. Twenty public assets do not establish quality on a user's complete photo library, music collection, recorded meetings, or long movies.