We run our own support on QuickRabbit. Here is how it is built.
Md Nazmul Hosan ·
Documents, screenshots and videos in one index, a confidence score that decides when to hand over, and a rule that the AI never guesses.
Short answer: QuickRabbit indexes documents, screenshots and videos, answers with sources when it is confident, refuses off-topic requests, and opens and routes a ticket when it is not sure.
Door Soft sells POS, ERP and billing software. Most support questions are the same few dozen, asked in different words at all hours. The answers existed in PDFs, screenshots and setup videos, but nobody could find them fast enough. QuickRabbit is the AI help desk we built to fix that, and it answers our customers today.
One index for documents, screenshots and video
QuickRabbit reads PDFs, Word files, FAQs and articles; reads screenshots with OCR and a vision model to pick out error messages; and transcribes videos, extracting key frames. Everything is split into chunks, tagged with source, version and the roles allowed to see it, and stored in PostgreSQL with pgvector.
Answer, refuse or hand over
- Answer, citing the sources used.
- Refuse politely when the question is off topic.
- Hand over when unsure, when the customer is frustrated or urgent, or when they ask for a person. It writes the ticket, classifies it, sets priority and assigns the best agent with a summary attached.
The rule that matters most
The AI never guesses. A wrong answer delivered confidently costs more than a ticket, because the customer acts on it.
What we are measuring
Our design target is to resolve up to 90% of eligible questions without staff. That is a target, not a result. We are tracking resolution, first-response time and handoffs and will publish 60 days of real numbers on the QuickRabbit case study.
Share
Nazmul founded Door Soft in Dhaka and leads system design and delivery. He writes about building AI that works in production for mid-size companies, and about the products Door Soft runs itself.
About the author