# Module 5 practice · combine without overstating

Use [the lesson](../lessons/05-compare-semantic-hybrid-rerank.md), then compare [the solutions](../solutions/05-hybrid.md).

**A. Geometry with declared inputs.** Use query vector `(1,1)` and document vectors A=`(1,0)`, B=`(0,1)`, C=`(0,0)`. Calculate cosine similarity to each. Predict the order from `fixed_vector_ranking` when IDs are `a,b,c`. Why is this not evidence that a text encoder understands a question?

**B. Rank fusion.** For lists `[a,b]` and `[c,a]`, use constant 10 and compute each RRF score. Which document wins? What happens to a relevant document `d` that appears in neither list? Verify using `reciprocal_rank_fusion`.

**C. Constraint boundaries.** A passage contains `Books may be borrowed.` Predict what the explicit required term filters `books`, `book` and `laptop` do. Explain two ways a manual exact-term filter can reject a useful passage. Is this implementation a BERT reranker?

**D. Write an honest comparison proposal.** Describe a future lexical-versus-dense experiment in six sentences or fewer. Name the frozen sources, labels, model/preprocessing information, common cutoff, per-case review and what counts as a held-out question. State that no neural comparison has run in this course package.

Self-review: correct cosine/RRF arithmetic, no invented candidates, explicit exact-token limitations and honest experiment status.