Fairytale Hunt: In-Browser embedding search text game (www.ghostweather.com)

Fairytale Hunt is an interactive in-browser demo that turns Project Gutenberg fairy-tale sentences into an embedding-driven search game: highlight a phrase with your mouse or finger and the app finds similar passages from other tales, scoring hits by rarity across curated categories (e.g., magical creatures). It’s both a playful UI and a technical proof-of-concept showing you can build client-side semantic search and scoring without a server: embeddings are computed with the tiny bge-micro model via Transformers.js, similarity search and indexing run in the browser using the client-vector-search Node package, and the dataset was sentence-tokenized and filtered with a small custom spaCy model for “descriptiveness.” For AI/ML practitioners this demo highlights practical trade-offs for lightweight, privacy-friendly retrieval: small embedding models and in-browser indexing make instant, offline semantic search feasible for demos and prototypes, but the dataset was drastically reduced to four authors and still contains duplicates and potentially offensive historical content. The scoring pipeline combines unigram/bigram seeds, a Python weighting script, and iterative adjustments using Claude; icons were generated with glif.ai via a Flux LoRA. The repo (github.com/arnicas/simple-embedded-text-navigator) provides code and artifacts for engineers who want a compact stack for client-side NLP exploration or educational tooling.