Improve Accuracy Over Time

A Nyckel prediction endpoint is not a static model. Every new labeled example makes the model smarter. Here is how to improve accuracy systematically.

Start small, iterate fast

You do not need a large dataset to start. Begin with 5–10 examples per label to get an initial model. Invoke it, review the predictions, and correct mistakes. Repeat.

The first few rounds of correction typically produce the largest accuracy gains.

Prioritize corrections over new data collection

Fixing a wrong prediction is often more valuable than adding a new example from scratch, because it comes from a case the model actually got wrong.

The place to do this is the Review tab in the console. Invoke capture has already set aside the predictions most worth your time — low-confidence predictions, a random slice of traffic — so annotation effort goes where it moves the model most. Confirm the right predictions, fix the wrong ones; each action becomes a training sample and Nyckel retrains automatically.

If your application learns true labels on its own — a support agent resolves the ticket, a user flags a wrong answer — you can also feed those errors back as labeled samples through the API and manage sample selection yourself. See Build a feedback loop.

Focus corrections where confidence is lowest

Low-confidence predictions are where the model is most uncertain. Correcting these teaches the model the most. See Confidence scores and thresholds.

Add examples for underperforming labels

If one label consistently has lower accuracy than others, add more labeled examples for it. Look at the accuracy breakdown per label in the Nyckel console.

Avoid adding noise

Do not add examples that are mislabeled, ambiguous, or inconsistent. Noisy training data reduces accuracy. When in doubt, leave an example unlabeled rather than guessing.