Orin Studio
Handcraft brand × AI customer service
80 customer messages a day — nine in ten answered before a human sees them
Speculative scenario: Orin Studio is a Taiwan handcraft lifestyle brand selling across Shopee and its own storefront. As orders grow, the LINE official account receives 80+ messages daily — 90% of them the same questions: in stock? convenience-store pickup? how long to deliver? custom embroidery? Two staff rotate through copy-pasting answers, burning time on predictable questions that don't need a human at all.
RAG (Retrieval-Augmented Generation) with Claude Haiku lets the bot query the brand-uploaded knowledge base — product catalogue, FAQ, return policy, shipping rates — before answering. Routine questions handled instantly by AI; for anything the bot is not confident about (custom order details, anomalous requests), a smooth handoff to a human agent — no hard wrong answer, no silent wait. The owner updates the knowledge base through the admin panel — change the return policy today, and the bot quotes the new version in the next message.
The bot knows what you teach it — nothing more, nothing less
Upload product FAQs, catalogue PDFs, return policy — the system chunks and vectorises automatically. Before answering, the bot retrieves from your knowledge base. It cites your content, not model hallucinations. Update a policy, add a product, change shipping rates — update the document, the bot syncs immediately, no retraining, no engineer. Today: 87 conversations, 94.3% resolved by AI, 5 handed to a human.
RAG, not fine-tuning
Fine-tuning requires large datasets, takes time, and costs money. RAG lets the model query the knowledge base at inference time — no retraining on updates, and higher accuracy because the bot cites source documents rather than "remembering" from training data.
When unsure, hand off — don't bluff
When confidence falls below a threshold, the bot proactively offers a human handoff instead of hedging an answer. Brand trust matters more than reply speed — customers can wait 5 minutes for a human, but will not accept a bot giving wrong refund information.
Claude Haiku, not GPT-4
The bottleneck for a support bot is latency and cost, not intelligence. Haiku responds 4–6× faster than large models on FAQ-retrieval tasks, at 80% lower cost, with no accuracy loss when knowledge base context is available. Save the large model for tasks that actually need reasoning.
What this build would use
- Next.js 15
- Claude Haiku (Anthropic)
- RAG / pgvector
- Supabase (PostgreSQL)
- LINE Messaging API
- Resend
- Vercel