Few-shot text classifier

About this tool

few-shot-text-classifier labels text from a small set of examples that you provide at run time. Paste one labeled example per line, such as billing,invoice charge refund, then paste the document or list of lines you want classified. The tool builds a deterministic local similarity model from that support set and returns the predicted label, confidence, per-label scores, shared terms, and nearest example.

This is useful for quick triage and lightweight labeling: support tickets, feedback themes, short content moderation queues, routing emails into buckets, or testing whether a proposed label taxonomy is separable before you train a larger model. It does not call an API, download embeddings, or learn a persistent model. Every run uses only the examples in the Labeled examples box.

Worked example

With these examples:

billing,invoice charge refund
billing,subscription payment failed
support,password reset login issue
support,account locked cannot sign in
sales,pricing quote enterprise plan
sales,demo request for buying team

and this text:

I cannot sign in after resetting my password.

the default centroid + cosine + TF-IDF settings predict support, show the support score above the other labels, and list terms such as password and sign as the explanation. Switch Classify text as to One document per non-blank line to batch-label a list and choose CSV when you want one result row per input line.

Options and limits

• Labeled examples must contain at least two distinct labels. Each non-empty line is split into label and text using tab, comma, pipe, or colon; Auto chooses the separator that appears on the most lines. Lines beginning with # are ignored. The examples input is capped at 1 MiB, 5,000 examples, and 200 labels.
• Decision method controls how label scores are built. Label centroids averages each label's examples and is the steadier default. k nearest examples lets the closest k examples vote by similarity. Best single example lets one very close example win for its label.
• Similarity metric can be cosine, dot product, inverted Euclidean distance, or Jaccard term overlap. Cosine is usually best for text of uneven length. Jaccard compares term sets only, so it ignores the weighting option.
• Feature weighting can be TF-IDF, raw term frequency, or binary presence. TF-IDF downweights words that appear across many examples; binary helps when every input is very short.
• Feature analyzer can use word n-grams or character n-grams. Character n-grams with length 3-5 are handy for typos, short names, and languages that do not use spaces between words.
• Lowercase, Strip accents, Remove English stop words, Sublinear term frequency, and Minimum example frequency tune how the vocabulary is prepared before scoring.
• Minimum confidence reports uncertain instead of the top label when the winner's vote share is below the threshold. Confidence is a relative vote share across labels, not a calibrated probability.
• Labels to show controls the score table; 0 lists every label. Show explanation terms adds the strongest shared terms and nearest example.
• Text to classify is capped at 256 KiB. Batch mode classifies up to 1,000 non-blank lines.
• This is a lexical similarity classifier, not a semantic embedding model. It will not infer that refund and reimbursement match unless similar words appear in your examples. Add representative examples, use character n-grams for typo tolerance, or lower the confidence threshold when labels are intentionally broad.

FAQ