AI0-001 Implementing AI Solutions Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Manhattan distance

  Why it's wrong here

  Manhattan distance sums absolute coordinate differences, suiting grid-like or sparse feature spaces rather than the angular closeness that text embeddings encode. It is tempting because it is a valid distance metric and cheaper than Euclidean, and it would be correct for high-dimensional sparse data where L1 geometry matches the domain.

• ✗

  Euclidean distance

  Why it's wrong here

  Euclidean distance measures the straight-line magnitude between vectors, but in high-dimensional embedding spaces typical of RAG systems, this metric fails to capture directional similarity, which is critical for matching semantic meaning. Cosine similarity is preferred because it normalises for vector length, focusing purely on angle. Euclidean distance is tempting because it is intuitive for low-dimensional spatial tasks, such as clustering physical coordinates or nearest-neighbour searches in geometric data.

• ✗

  Dot product

  Why it's wrong here

  Dot product is unbounded and scales with vector magnitude, so it ranks longer or larger-norm chunks above genuinely similar ones unless embeddings are normalised. It is tempting because for unit-normalised vectors dot product equals cosine similarity, and it would be correct in a store that guarantees normalisation.

• ✓

  Cosine similarity

  Why this is correct

  Cosine similarity measures the angle between query and chunk embeddings, ignoring magnitude, which suits text embeddings where direction encodes semantic meaning. It is the default metric in most vector stores such as Azure AI Search.