Business Context
BrightDesk, a SaaS helpdesk platform, wants to launch a GenAI chatbot for customer service across web chat and email deflection. Before choosing a model or building a RAG pipeline, the team wants to identify the first three technical constraints that will determine whether the solution is feasible and safe in production.
Data
- Volume: ~2.5M historical support tickets, 180K help-center articles, and 12K internal policy documents
- Text length: customer messages range from 5-1,200 words; knowledge documents range from short FAQs to 20-page policy pages
- Language: English (88%), Spanish (9%), French (3%)
- Label distribution: historical tickets include issue type, escalation outcome, and CSAT, but no direct labels for answer quality or hallucination
- Noise: duplicated macros, outdated articles, HTML fragments, signatures, PII, and inconsistent formatting
Success Criteria
A strong answer should prioritize the three constraints that most affect architecture and rollout, explain how to measure them, and propose a practical investigation plan. “Good enough” means the chatbot can answer common support questions accurately, cite grounded sources, and escalate safely when confidence is low.
Constraints
- P95 response latency must be < 2.5 seconds
- Customer data cannot leave the company’s VPC
- Answers must be grounded in approved support content
- The system must support multilingual queries
- Budget target is <$0.015 per resolved conversation turn
Requirements
- Identify the first three technical constraints you would investigate.
- Explain why each constraint matters for a GenAI customer-service chatbot.
- Describe the data analysis or experiments you would run first.
- Propose an initial NLP pipeline, including preprocessing and retrieval/generation components.
- Define how you would evaluate whether the constraints are acceptable for launch.
