Classification

Label text against your own labels in one call — ticket triage, moderation, lead scoring.

POST /v1/classify assigns your labels to text — support-ticket triage, content moderation, lead qualification, intent detection — without prompt engineering. The gateway builds a strict schema from your labels, so the model can only ever answer with one of them (out-of-label answers are caught and retried server-side).

curl https://api.gopuram.net/v1/classify \
  -H "Authorization: Bearer $GOPURAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "my card was charged twice this month and I want it fixed",
    "labels": ["billing", "technical-support", "sales", "feedback"]
  }'

{
  "object": "classification",
  "model": "google/gemini-3.6-flash",
  "label": "billing",
  "confidence": 0.92,
  "usage": { "cost": 0.00021, "steps": [ … ] }
}

Request

Field	Type	Notes
`input`	string	up to 100k characters
`labels`	string[]	2–50 unique labels, ≤200 chars each
`multi_label`	boolean	`true` → returns `labels: [{label, confidence}, …]`, every label the model asserts, sorted by confidence
`model`	string	optional structured-output chat model override

About confidence: it is the model's self-estimate, useful as an ordinal signal (0.9 beats 0.5) but not a calibrated probability — our upstreams don't expose token logprobs uniformly across 250+ models, and we won't dress up a guess as statistics. For routing decisions, thresholding at ~0.7 works well in practice; for anything regulatory, keep a human in the loop.

Batching: send one request per item — requests are cheap (a classification costs fractions of a cent) and parallelize cleanly client-side.