Configure the learning session
Change datasets, modes, thresholds, difficulty, timing, and accessibility without reloading.
Map the phrase to its intent
Drag the message card onto an intent node, or click a numbered node.
Classify a user message
Intent library
Select, create, rename, or remove intents.
Entity recognition workbench
Test built-in patterns and add custom regular-expression rules.
Custom entity rule
Rules are stored with the current game dataset.
Live chatbot test
Try unseen wording, inspect confidence, and correct mistakes.
Prediction inspector
Review the top candidates and classifier evidence.
Performance analytics
Plotly.js visualises confidence, coverage, accuracy, errors, and dataset balance.
Achievements and history
Complete milestones to unlock badges, stars, and a training certificate.
Recent predictions
Save, import, and export
Browser storage keeps progress without a database.
1. Define clear intents
Each intent should represent one user goal. Similar goals need distinct examples and descriptions.
2. Add varied phrases
Use synonyms, polite wording, short requests, misspellings, and natural sentence structures.
3. Balance the dataset
Large differences between intent example counts can bias a simple classifier toward common intents.
4. Label useful entities
Entities capture variable details such as dates, products, locations, order numbers, and names.
5. Test unseen wording
Use messages absent from training data. Correct weak predictions by adding representative examples.
6. Tune confidence
A higher threshold reduces risky replies. A lower threshold reduces fallback frequency.
7. Study confusion
Repeated confusion between two intents suggests overlap, weak wording, or missing training phrases.
8. Write safe fallbacks
Fallbacks should admit uncertainty, request clarification, and guide users toward supported topics.
9. Improve continuously
Review failed messages, retrain, compare metrics, and preserve successful dataset versions.