Custom labels at request time
No training step. Provide any set of labels and the model scores each against the input.
Zero-Shot Classification
Classify text into custom labels at request time — no training required
You send a single `text` and a list of candidate `labels`. The service scores each label against the text using a Natural Language Inference model and returns the top label, its score, and an `all_scores` map of every label → score. No training step. Synchronous. Note: the response uses an `all_scores` object (label → score), not an array of `[label, score]` pairs. Earlier docs were wrong.
textanalysisnlp
Overview
Features
Zero-shot via NLI
For each candidate label a hypothesis is constructed and scored as entailment of the text.
Single best + all scores
Response includes the top label and a complete label→score map for the rest.
Multilingual
Works in any language the underlying NLI model supports — keep labels in the same language as the text for best results.
Use Cases
Ticket routing
Classify support tickets into departments (billing / technical / sales) without training a custom model.
Content tagging
Assign custom tags to articles, reviews, or posts from an arbitrary label set.
Intent detection
Detect user intent (greeting, question, complaint, request) from a label list.
Input / Output
Input
text: string + labels: array of strings
JSON body
Output
label, score, all_scores (label → score map)
JSON
Specs
- Latency
- ~1-3 s (scales with number of labels)
- Async
- false
- Rate Limit
- Per API key
- Max Input
- Model-dependent, typically 512 tokens for the text
Quickstart
Prerequisites
- -A CN8 Gateway API key with text-classify in allowed_services
1. Classify with custom labels
text-classifyPOST text + a labels array. The service returns the best label, its score, and the score for every candidate.
POST/v1/proxy/text-classify
{
"text": "This new phone has amazing battery life and a great camera.",
"labels": ["technology", "food", "sports"]
}Response
{
"status": "success",
"data": {
"label": "technology",
"score": 0.7642,
"all_scores": {
"technology": 0.7642,
"sports": 0.0023,
"food": 0.0001
}
}
}Use the same language for labels as the input text. all_scores is an object (dict), not an array of pairs.
Services
View in API ReferenceZero-Shot Classification
POSTsyncScore user-defined labels against a text via NLI. Returns top label + an all_scores map.
/v1/proxy/text-classify
Pricing
Billed per request, independent of number of labels.
| Service | Unit | Price |
|---|---|---|
| Zero-Shot Classification | item | $0.006/request |
Guides & Tips
Important Notes (verified against the live upstream)
- -Single text per call — the request body is `text` (string), not a `texts` array.
- -`all_scores` is a dict, not a list of `[label, score]` pairs. Earlier docs were wrong.
- -Optional flags `label_descriptions`, `multi_label`, `threshold` are silently
- -accepted but do not change the response shape. They appear to have no effect
- -on the live upstream — treat them as best-effort.
- -Cost is per request, not per label.
How it works
- -For each candidate label a hypothesis like "This example is about X" is built.
- -The model scores how well the text entails each hypothesis using NLI.
- -The top label is exposed at `data.label`; the full ranking lives in `all_scores`.
Tips for best results
- -Keep the label set small (3–20 labels). More labels means more inference time.
- -Use labels in the same language as the input text.
- -For nuanced taxonomies with example texts per class, use Few-Shot Classification.
FAQ
Q: How do I rank all labels?
A: Sort the all_scores object by its values (descending). The top entry equals data.label.
Q: Can I do multi-label?
A: The flag is accepted but the response shape is the same — apply your own threshold to all_scores.
Q: What if I have example texts per category?
A: Use Few-Shot Classification — it embeds your examples and computes class prototypes.
Related Products
Changelog
1.2 (2026-04-29)
- -Aligned to live upstream: response is flat `{label, score, all_scores}` (no results[] wrapper).
- -Fixed `all_scores` shape: it is an object, not a list of [label, score] pairs.
- -Documented that label_descriptions / multi_label / threshold flags have no observable effect today.
1.1 (2026-02-23)
- -Aligned with zero_shot_classifier prose (now superseded).
1.0 (2026-01-26)
- -Initial catalog.