A sales-operations assistant is asked questions like "what was our average deal size in EMEA last quarter, by product line?" The CRM opportunity records were exported, embedded and added to the vector index, and answers are plausible but numerically wrong. What is the root cause and the best fix?

Answer: C — Similarity search can't aggregate over records; give the assistant a validated query tool against the CRM data

Aggregates need computation over all matching records. Vector retrieval returns a handful of similar-looking rows, so the model estimates from a sample. A structured, validated query tool over the source data returns the correct computed result. (a) and (b) improve which rows come back but still return a sample instead of a computation. (d) reasons harder over incomplete data.