Harvey Engine
Internal Guide & Rules
Welcome to QA Testing
The Harvey Engine is trained through the interactions and feedback from the Vind'r team. Please read these best practices carefully before you start testing.
⚠️ Privacy & Data (Crucial)
Although our API environment is secured and Google does not use this data for public model training, it is a strict rule that we upload no Personally Identifiable Information (PII) during the beta phase.
✓ DO Upload
- • Translated Amazon Notifications
- • Account Health warnings
- • Product ASINs and Titles
- • Screenshots of the "Policy Compliance" page (without names)
- • Manufacturer documentation (CE, SDS, ingredients)
✕ DO NOT Upload
- • Passports or ID cards of customers
- • Full corporate bank statements
- • Names or addresses of consumers
- • Internal, unencrypted passwords
💡 Best Practices: How to prompt Harvey?
The rule of thumb for Harvey is: Garbage in, Garbage out. If you give him half a sentence, he will make assumptions or reject your request via the internal 'Gatekeeper'. Provide him with context.
1. The Short Complaint (Bad) vs. The Context (Good)
Harvey will reject this because he doesn't know the Amazon rule (the root cause).
2. Handling Screenshots
Harvey is excellent at reading images (OCR). Make it easy for him:
🔄 Why Are We Testing? (The Feedback Loop)
Below each of Harvey's answers, you'll see a 👍 and a 👎 button. This is the most important part of the testing phase!
When you click a thumbs up/down, your original prompt AND Harvey's response are safely stored in our database along with a score (+1 or -1).
At the end of the week, we review which answers received a 'thumbs down'. Was the tone too aggressive? Did he miss an Amazon Policy? We then adjust the core prompt so he never makes that mistake again.
Your feedback trains the model