async def handle_guest_message(msg):

1. Intent classification (small model, via unified gateway)

intent = await ai_gateway.chat(

model="cheap-classifier",

prompt=f"分类以下民宿咨询: {msg.text}",

tag="intent-classify"

)

2. Non-inquiry types go directly to rules

if intent in ("complaint", "refund"):

return transfer_to_host(msg, reason=intent)

3. Assemble controlled context

context = build_listing_context(msg.listing_id) # listing factual data

4. Generate reply (large model, hard three-sentence constraint)

reply = await ai_gateway.chat(

model="main-model",

prompt=REPLY_TEMPLATE.format(context=context, question=msg.text),

tag="auto-reply",

max_tokens=150

)

5. Confidence safety net

if reply.confidence < 0.7:

return transfer_to_host(msg, reason="low-confidence")

return send_to_guest(msg, reply)

## Final Thoughts

The biggest lesson from this refactoring: **most AI application failures stem not from the model, but from the engineering surrounding it**. Get context, boundaries, and the access layer right, and an ordinary model can do good work; get them wrong, and even the most powerful model will fabricate, spiral out of control, and have you crawling out of bed at midnight to fix code.

If you're building a similar AI application, my advice is to build your access layer on a unified gateway from day one—don't be like me, running around with a naked API key for two weeks before waking up. You can start here: https://api.thistoken.ai/register

---

Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.