The interesting result in this household-chatbot writeup isn’t the chatbot — it’s the query router. The part that decides which metadata slice of the vector index to search runs fine on a 0.6B model once you fine-tune it. Most RAG stacks spend a frontier model on this and never notice.

• 🎯 Routing is classification, not generation. Mapping “when did we replace the pool pump?” to pool before retrieval, to shrink the candidate set, is intent classification wearing a RAG hat.
• ⚠️ Prompting a tiny model fails hard. Qwen 3:0.6B scored 10% (13/131) zero-shot — overusing broad labels and inventing categories outside the allowed list.
• ✅ Fine-tuning got it usable. Unsloth + QLoRA on ~850 examples took accuracy to 79%, though it still emitted fragments like ac/air instead of hvac.
• 💡 Output schema did the rest. Remapping categories to opaque two-letter codes with no semantic overlap pushed it to 92%. The fix wasn’t more data — it was removing the model’s room to be “almost right.”
• 🔍 The residual errors are ontology, not noise. Tankless water heater keeps landing in pool. No fine-tune papers over a labeling problem.

I keep seeing teams reach for the big model on the intent layer because it’s already in the stack. This is the counter-argument: a 600M classifier you run locally, evaluate against a 131-case battery, and retrain on user feedback. The HN discussion argues over where that line sits.

If a two-character code buys 13 points over semantic labels, how much of your own eval gap is the model versus the shape of what you’re asking it to emit?