SliSearch help
How search works
Last verified: 2026-09-02
SliSearch does not host models. It reads what marketplaces publish, stores the description of each listing, and sends you to the marketplace to download. This page says how a query becomes a page of results.
What is indexed#
For each listing SliSearch keeps the public metadata: the title, the creator, the tags, the category, the file types, the licence as the marketplace states it, the publication date, and the link.
It caches one preview image on its own origin, so a results page does not make your browser call five marketplaces.
It stores no model files. There is nothing to download here, and every result opens on the marketplace that published it.
How a query matches#
A query is matched three ways at once, because one way is never enough:
| Way | What it catches |
|---|---|
| Full text | The words of the title and description, stemmed |
| Title similarity | A near miss, a typo, a different word ending |
| Tags | A word the creator used as a label but never in prose |
The three run as three separate indexed lookups rather than one condition with
OR in it. Written as one condition the database cannot use an index for each
part, and the query goes from thousands of rows examined to hundreds of
thousands.
How results are ranked#
Every matching listing gets a score. The weights are fixed, and they are these:
| Weight | Signal |
|---|---|
| 0.50 | Full text rank against the query |
| 0.15 | Similarity between the query and the title |
| 0.15 | Likes, damped |
| 0.10 | An exact tag match |
| 0.10 | Freshness, decaying over one year |
A listing whose link is known to be broken keeps its score multiplied by 0.25. It is pushed down rather than removed, because a broken link is sometimes a marketplace being briefly unavailable rather than a listing that is gone.
Likes are damped rather than counted. A listing with 10,000 likes is not a thousand times more relevant than one with 10, and an undamped count would mean the same few famous models answered every query.
NOTE
Relevance needs a query to mean anything. With an empty search box every listing scores the same, so the ranking falls back to the most liked, which is what browsing with no query is actually asking for.
No source owns the page#
Ranking by score alone lets one marketplace fill an entire page, and it will: a source with more listings and more likes wins every comparison.
So each listing is also given its rank within its own source, and listings are grouped into bands of that rank. Band one holds each source's best few, band two the next few. Results are ordered by band first and by score inside it, so the strongest results still come first while no source can supply more than a band before another one appears.
Nothing is dropped by this. The result count and deep pagination are exactly what they would otherwise be.
Sorting#
| Sort | Order |
|---|---|
| Best match | The score above. Falls back to most liked with no query |
| Newest | Publication date, then when it was last seen |
| Most liked | The like score |
| Most popular | The popularity score, then likes |
The source diversity bands apply to every sort, not only to best match.
What is excluded#
- A source with its kill switch engaged returns nothing, and the page says so rather than quietly returning fewer results.
- Safe search is on by default.
- A listing withdrawn at the marketplace stops appearing, and one that was reported and removed here shows as unavailable rather than disappearing silently.
Where to look next#
- Licences, as reported, for how a marketplace's own wording becomes a filter.
- Reports, for what happens when you report a listing.