​
General Usage Tips

  • If you experience that it creates too much and goes into a loop, try to limit the number of questions by simply asking it to limit the number of questions

​
Wizflow Assistant Training Lab

​
Assistant Feedbacks

​
What It Does

Collects user feedback on AI assistant responses to improve our wizflow assistant user experience. Key Benefits:
  • Captures user satisfaction and detailed issues
  • Creates training datasets for model improvement
  • Provides insights into assistant performance
  • Measure User Satisfaction
After generation completes, users see thumbs up/down buttons:
  • 👍 Thumbs Up: Good response, no further action needed
  • 👎 Thumbs Down: Opens detailed feedback form, allow user to specify what went wrong with generation.
Feedback

​
Feedback browser and editor

Only users with elevated permissions can view the training lab.
View all collected feedbacks in training lab tables.

​
Feedback table

Shows all collected feedback with:
  • User prompts and comments
  • Rating (Satisfactory/Unsatisfactory)
  • Issue tags
  • Attachment count
  • Review status
  • Actions (View | Delete)
WizflowAssistantLab

​
Feedback editor

Each feedback has three states accessible at feedback editor, which can easily accessed from feedback table action column
Feedback states
https://github.com/user-attachments/assets/e5c87799-9a86-457a-9a65-53e9010d4318
  • Original: Template before AI changes
  • Generated: AI-modified template
  • Ideal: Corrected response (created by reviewers)
Overall collection and dataset generation process
  1. Users submit feedback
  2. Internal reviewers create “Ideal” responses
  3. Export script generates training datasets
  4. Data used for SFT and DPO training

​
Technical Overview

​
Data storage

All feedback stored in ChatflowAssistantFeedback table including:
  • Complete template states
  • User context and attachments
  • Ratings and comments
  • Issue categorization

​
Fine-tuning

  • SFT (Supervised Fine-Tuning): Transform feedbacks collected into single use samples
  • DPO (Direct Preference Optimization): Use rating to create dataaset consisting of preferred and non-preffered output for a user prompt. This is a different kind of optimisation that helps in improving model accuracy by comparison.

​
Key features

  • Optional feedback submission
  • Binary rating system
  • Predefined issue tags (user-assigned)
  • Role-based access control
  • Automated export functionality via script

​
Security & Privacy

  • Anonymous feedback collection
  • Permission-based training lab access
  • No personal user data captured
  • Secure training data exports

​
Next Steps