Specializes in retrieval-augmented systems, chunking, retrieval, reranking, vector DBs, RAG eval.
Deep dive into embedding models, similarity metrics, and vector database internals (HNSW, IVF).
By the end of this week, you can pick HNSW vs IVF for a given workload and explain the recall versus latency trade-off cold.
RAG roles screen hard on vector-DB internals. Knowing HNSW versus IVF trade-offs and recall/latency curves is table stakes.
Watch out: Mixing up cosine similarity with raw dot product on un-normalized vectors is the single most common interview slip.
Master chunking strategies, hybrid retrieval (BM25 + dense), re-ranking, and context management.
By the end of this week, you can debug a 'wrong chunk retrieved' incident step by step on a whiteboard.
Chunking is where most RAG systems fail silently. Interviewers ask how you'd debug a system that retrieves the wrong chunk.
Watch out: Skipping re-ranking and ignoring BM25+dense hybrids gets caught. Senior loops expect you to know when each helps.
Build production RAG: ingestion pipelines, citation tracking, failure modes, and evaluation.
By the end of this week, you can design a full production RAG pipeline with ingestion, citations, and a graceful-degradation story.
Citation tracking and graceful-degradation paths are the bar for senior RAG roles. They reveal whether you've shipped production.
Watch out: Forgetting citation tracking is the fastest way to lose a senior RAG interview. Make it a default, not an afterthought.
Advanced patterns: multi-hop retrieval, agentic RAG, caching, serving, and observability.
By the end of this week, you can argue when multi-hop or agentic RAG actually pays for itself versus when it just adds latency for no gain.
Top RAG roles want multi-hop and agentic patterns plus a story about caching and cost. This separates IC-3 from staff.
Watch out: Reaching for agentic RAG on every problem signals you haven't read the cost numbers. Default to simple; reach for agentic with a reason.