Business Context

BrightDesk, a B2B SaaS company, wants a proof-of-concept GenAI application that answers customer questions using internal product documentation, support articles, and past ticket resolutions. The client does not want a slide deck; they want a working NLP prototype that can be evaluated by support managers within 4 weeks.

Data

You have access to approximately 180,000 English documents: 25,000 product docs, 90,000 help-center articles, and 65,000 historical support tickets with agent-written resolutions. Documents range from 30 to 2,500 words, contain markdown, HTML fragments, tables, code snippets, and duplicated content across versions. Roughly 15% of tickets are low-quality or incomplete. User queries are short natural-language questions, typically 5-40 words, with ambiguous product terminology and acronyms.

Success Criteria

A good proof of concept should answer at least 80% of benchmark questions with grounded, relevant responses, cite supporting passages, and keep median end-to-end latency under 2.5 seconds. The system should be easy to extend into production if the client approves the pilot.

Constraints

• No fine-tuning on proprietary foundation models during the POC
• Must run in the client’s cloud environment
• Sensitive ticket data must be masked before indexing
• Budget is limited to one small GPU or CPU-only fallback

Requirements

1. Design an NLP pipeline for a retrieval-augmented GenAI assistant.
2. Define preprocessing for noisy enterprise text, including chunking and deduplication.
3. Implement a baseline using modern Python libraries for embedding, retrieval, and generation.
4. Explain how you would evaluate answer quality, grounding, and failure modes.
5. Describe trade-offs between a lightweight POC and a production-ready architecture.
6. Specify what artifacts you would deliver to the client at the end of the POC.