Customer Feedback Classification: One Theme, a Priority Score, and Why Sentiment Is Not Enough

The decision is two questions, not one score. First: what is this comment mainly about? Second: how soon should a product person read it? Sentiment answers a different question — positive, neutral, or negative — and a weekly theme report disappears if you only store polarity. Support triage answers yet another question: which live queue should reply. A comment can be negative and still be a feature request, or positive and still describe a broken export. Opening a ticket for every comment also fills the queue with praise. Use sentiment-analysis-api when the dashboard needs polarity. Use support-triage or zendesk-ai-ticket-classification when someone must reply. Use content-moderation when the action is hide, ban, or escalate for safety. This page owns the offline corpus: NPS open answers, in-app feedback, public reviews, and survey free text that product teams bucket into a fixed theme list.

Classify one open-ended answer at a time. If a survey has three free-text boxes, make three rows so themes do not blend. Build a JSON state with the respondent’s own words plus a few fields you already trust: channel (nps, in-app, review), the numeric score if any, plan, locale, and product area when the form asked for it. Strip the boilerplate the form adds, such as a label “Additional comments:”. If the row arrived as email, trim quoted replies and signatures first — that work is email-classification — and only then classify the remaining comment. Do not paste screenshots, audio, or files. This API accepts text or JSON only; a written description of an image is not the same as showing the pixels. Anonymize names, emails, and account ids before the request leaves your system. The serialized state must be at most 8,000 characters, and the full validated request (state plus questions) must also fit in 8,000 characters, so leave room for the criteria. Trim to the respondent’s sentences. If another model summarized the comment first, record that fact on the row, because you are no longer classifying the original words. Request data is sent to the upstream provider; see privacy-and-history.

Most teams keep a thin rule layer for phrases they already trust, then a model for the rest. The table compares contracts and failure modes. It is not a benchmark, and none of the three replaces a labeled sample of your own comments.

The Playground ships an example named Feedback intelligence. It classifies a comment about a report export getting slower into a theme (performance, usability, feature, or billing) and a follow-up Score from low to very high. The request below keeps that comment and adds an explicit mixed theme so a short or multi-topic note is not forced into a product area. Question ids start with a letter and then use only lowercase letters, digits, and underscores, up to 40 characters. Instructions are 8 to 600 characters. Each criterion is 2 to 350 characters. Choice takes 2 to 12 themes. Score takes 2 to 10 ordered levels; list them in rubric order, here from low to very high. A Choice answer has choice, a probability per theme, and a confidence summary of that distribution. Confidence is not a measured accuracy rate. A Score answer has score, which can fall between levels, and that number is not a probability. Read both before any side effect. A theme is not permission to change the roadmap. Sign in, verify email, create a key under Console, API keys, and store it as JEVMODEL_API_KEY on the server. The call below is a live request and can spend allowance or tokens. Free verified accounts share 5 successful runs per UTC day across Playground, each Batch row, and this API, subject to site capacity. The fixed model is typesafe/jev-1.13 via OpenRouter. There is no idempotency key: if the response is lost, check history before you retry.

curl --fail-with-body https://jevmodel.app/api/v1/systemone \
  -H "Authorization: Bearer $JEVMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"state":{"channel":"nps","score":3,"comment":"After the latest update, exporting a report takes two minutes instead of ten seconds. We export every morning."},"questions":{"theme":{"type":"choice","instructions":"Classify the main theme of this customer feedback.","criteria":{"performance":"Performance or speed regression.","usability":"Confusing interaction.","feature":"Feature request.","billing":"Billing issue.","mixed":"Several themes, or not enough text to choose."}},"urgency":{"type":"score","instructions":"Rate product-team follow-up priority for this feedback.","criteria":["Low: optional improvement","Medium: limited friction","High: recurring workflow degraded","Very high: core workflow blocked for many users"]}}}'

Set three bands on your own held-out comments, not on a universal cutoff. High band: write the theme and the priority onto the row in your warehouse. That action is reversible. Mid-band, the mixed label, or two themes whose probabilities are close: show the comment to a reviewer with the top two themes attached, and let that person add a second theme by hand if the report needs it. Low band: leave the row untagged. Do not create a public roadmap item from a label alone. Calibrate which Score values should page someone; the score is a reading of your rubric, not a probability of being right. Re-check the bands when you rename criteria or the upstream model behavior changes. For a backlog, Console, then Batch, imports CSV, TXT, or JSONL up to 2 MB and 1,000 rows, applies one fixed question set, and exports row-level answers. Join those answers to your labels outside the site. Batch does not compute accuracy, and it is not an asynchronous bulk-job endpoint. A row is still one ordinary run. See confidence-and-abstention for thresholds and evaluation for the held-out split. If one comment must carry several themes at once, that contract is multi-label-text-classification, not a wider Choice.

Stay here when the input is an open-ended product comment and the output is one theme plus a priority. Leave for sentiment-analysis-api if you only need positive, neutral, or negative for a chart. Leave for support-triage if the text is a live ticket that needs a reply. Leave for email-classification if the raw input is still a mailbox thread. Leave for content-moderation if the action is a safety enforcement. Leave for multi-label-text-classification if the report must keep every applicable theme, not only the primary one. A Decision Flow can chain the theme question into a follow-up question, but running a flow is for paid plans only: subscribe or buy a token pack. Flows do not use the daily free runs. JevModel does not train a classifier on your comments, does not accept images or attachments, is not open source, and does not offer a weight download.