OPEN_SOURCE ↗
HN · HACKER_NEWS// 12d agoPRODUCT UPDATE
Marginalia Search ships NSFW filter
Marginalia Search has shipped an NSFW filter for its indie search engine, using LLM-labeled training data and a lightweight neural net so the classifier stays cheap enough to run on CPUs. The feature is live on the API now, with a UI toggle promised later.
// ANALYSIS
This is the right kind of boring: a moderation feature engineered for search latency and low base rates, not benchmark theater.
- –HN reaction was predictably split between curiosity and skepticism, with the maintainer framing the feature as useful for API consumers and anyone who wants safer defaults.
- –The build pipeline is the real story: qwen3.5 and ollama bootstrap labels, fasttext got tried and rejected, then Marginalia settled on a hand-featured single-hidden-layer net.
- –The post is unusually candid about false positives and the risk of over-filtering legitimate terms, which matters more than raw accuracy in a search product.
- –Shipping on the API first is a smart rollout, because NSFW moderation needs real traffic before a UI switch can be trusted.
- –For developers, this is a good reminder that sometimes the best production model is the smallest one that still behaves acceptably.
// TAGS
marginalia-searchsearchllmopen-sourcesafety
DISCOVERED
12d ago
2026-03-30
PUBLISHED
12d ago
2026-03-30
RELEVANCE
5/ 10
AUTHOR
speckx