--- title: "Keyword Extractor - Private AI Analysis | CanDoYa" description: "Extract keywords from text with private, on-device AI. Rank English words and phrases by semantic relevance, then copy or export results." url: https://candoya.com/tools/keyword-extractor/ locale: en type: browser-tool --- # Free AI Keyword Extractor Category: Social & Creator ## What is a keyword extractor? A keyword extractor finds the words and short phrases that best represent a text. This tool generates phrases from your English passage, embeds them with a local MiniLM model, and ranks them against the document's meaning. You get useful topic labels without sending the pasted text to an analysis server. **This extractor follows a KeyBERT-style method in your browser.** It builds one-to-three-word candidates from the passage, removes common English filler words, and uses cosine similarity between MiniLM embeddings to measure how closely each phrase matches the document. A small occurrence signal helps repeated terms without turning the ranking into a simple frequency list. The Apache-2.0 all-MiniLM-L6-v2 model maps text into 384-number vectors. Long documents are split into balanced chunks before their vectors are averaged, which avoids judging an entire article only from its opening. WebGPU uses roughly 30 MB of quantized weights when supported; the WebAssembly CPU fallback uses roughly 54 MB. Browser and model runtime files download separately and may be cached. ## How to use it 1. **Paste your text** - Enter an English article, report, transcript, or other topic-rich passage. 2. **Choose the detail** - Set the longest phrase to one, two, or three words and choose how many results you want. 3. **Extract keywords** - Run the on-device model and review relevance and occurrence counts for each phrase. 4. **Use the list** - Copy the phrases as a comma-separated list or download the scored results as CSV. ## Who it's for - **Content editors** can check whether a draft stays focused on its intended topic before publication. - **Researchers and students** can turn reports, lecture notes, or abstracts into a compact concept list. - **Knowledge-base teams** can draft tags for articles and support documents without uploading unpublished text. - **SEO specialists** can inspect topical coverage, then validate promising terms with a separate search-volume tool. ## Expert note MiniLM reads at most 256 word pieces per model input, so this implementation embeds up to ten document chunks and averages their normalized vectors instead of silently discarding everything after the opening section. ## Frequently asked questions ### Is my text uploaded to a server? No. Your passage is passed to a worker inside your browser, and model inference runs on your device. The worker downloads the public Transformers.js runtime and MiniLM model files from external hosts, but those requests do not contain the text you pasted. CanDoYa does not receive or store it. ### Is the keyword extractor free? Yes. You can extract, copy, and export keywords without an account or usage fee. Your browser supplies the computing power, so the first run can take longer while it downloads the AI runtime and model. Cached files often make later runs faster. ### How much text can I analyse? The tool analyses up to 8,000 Unicode characters per run and tells you if a longer passage was trimmed. It requires at least eight words. For a long document, analysing individual sections can reveal more specific phrases than processing the whole document at once. ### Does it extract single words or phrases? It can rank single words, phrases up to two words, or phrases up to three words. The phrase must occur in the source text. Increasing the phrase length can reveal precise topics such as urban forestry, but it also creates more candidates for the model to compare. ### How does AI keyword extraction work? The tool creates candidate phrases, converts the document chunks and each candidate into normalized MiniLM embeddings, then compares them with cosine similarity. It blends that semantic match with a small frequency score. This is closer to KeyBERT-style ranking than a plain word-count table. ### Does this show Google search volume or keyword difficulty? No. Relevance describes how well a phrase represents the pasted text, not how often people search for it or how hard it may be to rank. Use these results for topic discovery or tagging, then check search demand and competition in a dedicated SEO research product. ### Which languages does the model support? This page is intended for English. The MiniLM model was trained as an English sentence encoder, and candidate filtering uses an English stop-word list. Other Latin-script languages may return phrases, but their relevance scores are not calibrated or guaranteed to be useful. ### Will it work offline? It can work after your browser has cached the Transformers.js runtime, tokenizer, and model files, but cache availability is controlled by the browser. A first visit needs a network connection, and private browsing, storage cleanup, or cache eviction can require another download. ## Related tools - [Language Detector](https://candoya.com/tools/language-detector/) - [Named Entity Recognition](https://candoya.com/tools/named-entity-recognition/) - [Sentiment Analysis](https://candoya.com/tools/sentiment-analysis/) - [Parts of Speech Tagger](https://candoya.com/tools/parts-of-speech-tagger/) - [Text Summarizer](https://candoya.com/tools/text-summarizer/) --- *Markdown edition of , published for AI agents and other automated readers. Index of key pages: .*