v0.37.0: Your Chatbot Stops Making Things Up 📚
TLDR: A big AI accuracy release. Your chatbot now answers from your knowledge base reliably (with a new Train step), the fact checker catches wrong answers even on long replies, and your WhatsApp channels no longer get wrongly disconnected during Meta hiccups.
Here's what's new 👇
📚 Grounded Answers From Your Knowledge Base
The problem: Even when the correct answer (a price, a spec) was sitting in your knowledge base, your chatbot could still give a wrong one. It treated knowledge base info as lower priority than training data, lost track on follow up questions, and couldn't compare across products.
- Your chatbot now treats your knowledge base as an equal, trusted source, so the most specific answer wins instead of being overridden. 🎯
- It handles follow up questions ("how heavy is it?") and multi part questions correctly, pulling from the right product's full document.
- It can now answer "cheapest", "lightest", or "under RM X" style questions by comparing across your whole catalogue with exact numbers.
- New
Your existing chatbots keep working right away. To get the full benefit of the new comparison answers, give your knowledge base a Train run.
✅ More Accurate AI Fact Checking
The problem: On longer replies, the AI fact checker could quietly fail and send the original, uncorrected answer while logging it as a clean pass, so wrong facts still slipped through.
- The fact checker now reliably corrects wrong facts even on long replies, and it will never ship a reply it couldn't verify. 🐛
🔌 Channels Stop Disconnecting on Meta Hiccups
The problem: When Meta's servers were slow or timing out, ReplyLa could wrongly mark a healthy WhatsApp channel as disconnected, and it would then stay stuck (some channels were stuck for weeks).
- A temporary Meta timeout no longer disconnects your channel. It's only marked disconnected on a genuine login or token problem, and otherwise heals itself on the next check. 💪