Submit Feedback - Developer Platform

Submit Feedback

curl --request POST \
  --url https://api.trynawa.com/v1/feedback

POST

/

v1

/

feedback

Submit Feedback

curl --request POST \
  --url https://api.trynawa.com/v1/feedback

Tell NAWA when a classification was wrong, partially right, or correct. Feedback accumulates into pattern hints that NAGL injects on future classifications, so corrections compound over time.

This endpoint is free. It does not consume credits. Sandbox keys can submit feedback, and each submission counts against the 100-request sandbox lifetime cap.

Request

Headers

Header	Required	Description
`Authorization`	Yes	`Bearer nawa_live_sk_xxx` or `Bearer nawa_test_sk_xxx`
`Content-Type`	Yes	`application/json`

Body

Parameter	Type	Required	Description
`classification_id`	string	Yes	The `id` from the classification you are correcting (e.g. `cls_nw_lx4cqy2j2o1o` from the `/v1/classify` response).
`rating`	string	Yes	One of `correct`, `incorrect`, `partial`. How accurate the original classification was.
`corrected_intent`	string[]	No	The intent labels you believe the classification should have returned. Free-form strings; the standard labels are `question`, `complaint`, `praise`, `suggestion`, `spam`, `other`.
`corrected_sentiment`	string	No	One of `positive`, `negative`, `neutral`, `mixed`. What the sentiment should have been.
`comment`	string	No	Optional free-text explanation. Useful for dialect edge cases and intent ambiguity.

Example request

curl -X POST https://api.trynawa.com/v1/feedback \
  -H "Authorization: Bearer nawa_test_sk_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "classification_id": "cls_nw_lx4cqy2j2o1o",
    "rating": "partial",
    "corrected_intent": ["question", "suggestion"],
    "corrected_sentiment": "neutral",
    "comment": "The comment is a question AND a polite suggestion. The model caught only the question."
  }'

import { Nawa } from '@nawalabs/sdk'

const nawa = new Nawa({ apiKey: process.env.NAWA_API_KEY })

const { data, error } = await nawa.feedback.submit({
  classificationId: 'cls_nw_lx4cqy2j2o1o',
  rating: 'partial',
  correctedIntent: ['question', 'suggestion'],
  correctedSentiment: 'neutral',
  comment: 'The comment is a question AND a polite suggestion.',
})

from nawa import Nawa

nawa = Nawa(api_key="your_api_key")

result = nawa.feedback.submit(
    classification_id="cls_nw_lx4cqy2j2o1o",
    rating="partial",
    corrected_intent=["question", "suggestion"],
    corrected_sentiment="neutral",
    comment="The comment is a question AND a polite suggestion.",
)

Response

Success response (200)

{
  "result": {
    "id": "fb_nw_8k3r9vwxyz01",
    "object": "feedback",
    "classification_id": "cls_nw_lx4cqy2j2o1o",
    "rating": "partial",
    "acknowledged": true
  },
  "success": true,
  "errors": [],
  "request_id": "req_nw_fb01abc23def4"
}

Result fields

Field	Type	Description
`id`	string	Feedback record ID in `fb_nw_xxx` format
`object`	string	Always `"feedback"`
`classification_id`	string	Echoed back from the request
`rating`	string	Echoed back from the request
`acknowledged`	boolean	Always `true` on a 200 response. Confirms the feedback was persisted.

Response headers

Header	Description
`X-Request-Id`	Unique request identifier
`X-NAWA-Environment`	`sandbox` or `live`
`X-RateLimit-Limit`	Your tier’s per-minute request ceiling
`X-RateLimit-Remaining`	Requests remaining in the current one-minute window
`X-RateLimit-Reset`	RFC 3339 timestamp when the window resets

Error responses

Status	Type	When
400	`invalid_request_error`	Missing `classification_id` or `rating`; invalid `rating` / `corrected_sentiment` value; `corrected_intent` not an array of strings; malformed JSON body
401	`authentication_error`	Missing, malformed, revoked, or expired API key
429	`rate_limit_error`	Per-minute rate limit exceeded, or sandbox lifetime quota (100 requests) exhausted
500	`api_error`	Unexpected internal failure. Quote `request_id` to support.

See Errors for the full envelope shape and all error codes.

We especially value feedback on Arabic dialect edge cases and on comments where the same text expresses more than one intent. Both are harder for models than they look.