Build a Knowledge Base: Design Brief

You are designing a knowledge base for an AI agent that will run in a specific real-world context. Choose ONE scenario below, then design the knowledge base:

Scenarios (choose one)

A) Customer Support: A Refund & Returns Chatbot for an online retailer. Customers ask about returning items, getting refunds, and checking return status.

B) Internal Tool: An HR Policy Chatbot for a mid-size company. Employees ask about vacation time, parental leave, expense reimbursement, and benefits eligibility.

C) Educational: An Admissions FAQ Chatbot for a university. Applicants ask about application requirements, deadlines, scholarships, and deferral options.

D) Medical: A Symptom Screening Chatbot at a clinic. Patients answer screening questions and get guidance on whether to schedule an appointment.

Your Deliverable

Write a design brief for the knowledge base:

SCENARIO: [Your choice: A, B, C, or D]

KNOWLEDGE BASE DESIGN

1. **Authoritative source:** Where does each fact come from? (e.g., company policy doc, legal statute, doctor-approved guidelines). List 3–4 sources.

2. **What goes in:** List 5–6 key documents or topic areas that will be in the KB. Describe what each one covers.

3. **What stays out:** Name 2–3 things that will NOT be in the KB, even though users might ask about them. Explain why not.

4. **Ambiguities to resolve:** Identify 2–3 questions your knowledge base needs to answer clearly. For each one, write the exact wording that will go into the KB to avoid confusion.

5. **Retrieval risks:** Identify one query that your retrieval system might handle wrong (retrieves the wrong document, or misses a crucial one). How will you prevent it?

6. **Process:** Who owns the knowledge base? How often will it be reviewed or updated? What happens when a policy changes mid-year?

Evaluation

Your design is strong if it:

• Draws a clear line between what the agent will know and what it won't (no vague "everything relevant")
• Names the people and processes that will keep the KB accurate
• Anticipates at least one way retrieval or the agent could fail, and explains how you'd catch it
• Shows editorial judgment, not just technical implementation