Keyword Extractor
100% private — runs on your device, never uploaded. Works offline once loaded.
Paste text and get the phrases that best represent it, ranked by relevance using an on-device embedding model. Nothing you paste is uploaded.
Semantic keyword extraction
This tool uses a MiniLM sentence-embedding model (all-MiniLM-L6-v2) running on your device. It embeds the whole document and each candidate phrase into vectors, then ranks phrases by how close their meaning is to the document as a whole — the same idea behind KeyBERT.
Because it compares meaning rather than just counting word frequency, it surfaces phrases that genuinely represent the text, and it does it all locally so your content is never uploaded.
Why semantic beats frequency
A simple frequency counter rewards common words; semantic ranking rewards phrases that capture the document’s topic even if they appear only a few times. The tool considers one- and two-word candidates, removes stop-words, and drops redundant phrases already covered by a stronger one.
- SEO: find the themes a page actually covers
- Tagging: auto-suggest tags for articles or notes
- Research: see the main concepts in a document at a glance
Getting good keywords
Longer, well-written text produces better keywords, since the model has more context to compare against. For very short snippets there is little to rank. The relevance percentage is a similarity score, useful for comparing candidates rather than as an absolute measure.
Frequently asked questions
Is my text uploaded?
No — embedding and ranking happen entirely in your browser.
How is this different from a word-frequency counter?
It ranks phrases by semantic similarity to the whole text, not by raw counts, so it surfaces representative topics rather than just frequent words.
What does the relevance percentage mean?
It is the cosine similarity between the phrase and the document embedding — higher means a closer match. Use it to compare candidates.
Does it handle multi-word phrases?
Yes — it considers one- and two-word candidates and removes redundant overlaps.
Why is the first run slow?
A one-time model (~25 MB) downloads and caches on first use; it is one of the smaller AI models here.
Does it work offline?
Yes, once the model is cached.
Advertisement