Recherche sémantique (sur l'appareil)

Recherchez dans vos propres documents par le sens grâce à des embeddings embarqués : collez plusieurs textes, posez une question, obtenez un classement.

First use downloads the all-MiniLM-L6-v2 embedding model (roughly 25 MB) from Hugging Face. After that, everything runs on your device: your documents and query are never uploaded.

Important: this embedding model is English-centric. It ranks English documents well; Japanese and other non-English text usually matches poorly, because the model was trained mostly on English.

Documents (separate documents with a blank line; the first line of a block is used as its title)Query

Comment utiliser

Paste documents separated by blank lines (2–20). The first line of each block is shown as its title.
Type a query and press Search. The first run downloads the embedding model.
Documents are ranked by meaning, with cosine similarity scores.

Remarques

Embeddings are computed on-device with Xenova/all-MiniLM-L6-v2 via Transformers.js; neither your documents nor your query are uploaded. The model was trained mostly on English, so it is English-centric: non-English documents, including Japanese, often rank poorly. That limitation is stated inside the tool as well.

Questions fréquentes

How is this different from keyword search?: Keyword search only checks whether the same words appear. Semantic search converts text into vectors and compares meaning, so relevant documents can rank highly without sharing words. For exact-match needs, keyword search is still more reliable.
Do the scores mean anything absolute?: Treat scores as relative ranking signals within this document set. They change with the set, and 0.5 does not mean 'half correct'.