← American Express Interview Insights
They gave 30 minutes for this and I spent way too long on chunking strategy upfront.
Start by clarifying requirements and constraints (e.g., document types, scale, latency, compliance) to show adaptability. Then walk through the RAG pipeline stages—ingestion, retrieval, generation, evaluation, and operations—highlighting trade-offs and design choices at each stage. Emphasize how your design addresses American Express’s needs for accuracy, security, and scalability.
Pro tip: Anchor your answer in a concrete example or two (e.g., handling PDFs with tables, or mitigating hallucinations in financial Q&A) to demonstrate practical experience and maturity. Also, proactively mention compliance and data privacy considerations, which are critical in financial services.
Ask about document types, corpus size, update frequency, latency/throughput needs, and compliance requirements. This shows you can adapt to ambiguity and tailor the design.
Cover document parsing (e.g., PDF, HTML), chunking strategies (fixed-size, semantic), embedding model selection, and metadata extraction. Discuss handling of tables, images, and incremental updates.
Explain hybrid retrieval (dense + sparse), re-ranking, and query understanding. For generation, discuss prompt engineering, context window management, and model choice (e.g., fine-tuned vs. API-based).
Define offline metrics (retrieval recall, answer faithfulness) and online metrics (user feedback, latency). Describe A/B testing and continuous evaluation to catch drift.
Discuss scalability (vector DB sharding, caching), cost management, security (access control, encryption), and compliance (audit logs, data retention). Mention CI/CD for model updates.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.