Implemented here
Lowercase Latin-alphanumeric tokens, a fixed stopword list, normalized term frequency, smoothed inverse-document frequency and L2 normalization. Exact cosine distances feed DBSCAN with a deterministic input-order traversal. All processing stays in the browser.
What this cannot prove
No semantic embeddings, translation or trained intent labels. Synonyms and negation can fool word overlap. Noise is not a fault or priority score. Duplicate messages influence density; corpus changes alter IDF and clusters. Border points can depend on traversal order.
Proposed production path
Authorized, redacted ticket ingestion; multilingual embedding comparison; reviewer-labelled validation; cluster stability and drift checks; access controls. Only reviewed patterns become candidate chatbot or routing rules. No support reply is sent.