Module 2: Semantic Router¶
What you're building¶
Your agent currently answers everything, even off-topic questions like "Write me a Python script." In this module you'll add a semantic router that filters queries before they reach the AI.
You'll define two lists of example queries: an allow list for questions your agent should handle, and a deny list for ones it should block.
Setup¶
Enable the Semantic Router in your .env file:
Exercise¶
- Open
exercises/healthcare/semantic_router.pyin your IDE. - Add 2-3 more references to each route to improve classification. Each
Routehas:
| Field | What it does |
|---|---|
name |
"allow_list" lets queries through, "deny_list" blocks them |
references |
Example queries that represent this category |
distance_threshold |
How close a match must be (lower = stricter) |
Route(
name="allow_list",
references=[
"Show me my appointment calendar",
"When is my next appointment?",
"I need an update on my referral",
"Who is my primary care provider?",
"Is telehealth available for my visit?",
"What's my insurance status?",
"Can you help me?",
"What do you know about me?",
# Add 2-3 more healthcare queries...
],
distance_threshold=0.7,
),
Route(
name="deny_list",
references=[
"Write me a Python script",
"Tell me a joke",
"What's the weather like today?",
# Add 2-3 more off-topic queries...
],
distance_threshold=0.5,
),
Click for a hint
For the allow list, think about scheduling follow ups, rescheduling appointments, or greetings like "Hello". For the deny list, think about homework help, sports scores, or creative writing.
Verify¶
Stop the running server (Ctrl+C) and restart:
Open localhost:3040.
- Ask: "Do I have any upcoming appointments?" and click
Semantic-routerin the right activity panel to see it pass through - Ask: "Write me a Python script" and see it get blocked with a redirect message