Every term on this line is a follow-up question waiting to happen.
On your resume
Designed a secure enterprise knowledge platform with BM25 and vector retrieval, reciprocal rank fusion, reranking, citation-backed generation, and document-level RBAC and ABAC filtering before retrieval and generation.
How ready are you?
Cross-questions
5
A strong answer covers
Score = the sum over retrievers of 1 / (k + rank), with k usually 60.
It uses ranks, not raw scores, so BM25 and cosine scores never need normalizing.
The alternative is weighted score fusion after normalization.
Avoid
Not knowing what k does.
A strong answer covers
RBAC works on roles; ABAC on attributes such as department, region and classification.
Enforce access as a filter inside the retrieval query, and check again before showing citations.
Test for leakage with users from different groups.
Avoid
Enforcing access after generation.
A strong answer covers
Input screening, clear separation of instructions from retrieved content, and output checks.
Indirect injection hidden in documents is the hardest case, and no detector is complete.
So limit what the model can do: least privilege and no secrets in context.
Avoid
Claiming prompt injection is solved.
A strong answer covers
Retrieval: recall@k, MRR. Generation: faithfulness, relevance, citation accuracy. Safety: leakage and refusals. Plus latency and cost.
A golden set plus synthetic data from EvalSeed.
Avoid
One overall “accuracy” number.
A strong answer covers
On exact terms: IDs, product codes, names, acronyms and rare words. Vectors win on paraphrase; hybrid handles both.
Avoid
Saying vector search always wins.
Your numbers
Interviewers probe numbers first. Write exact values so you never have to improvise.