Similarity metrics and retrieval-specific metrics

This is Week 3, Lesson 4. The focus shifts from end-to-end quality metrics to component-level diagnostics. Specifically: how to measure whether the retrieval system is surfacing the right context before generation even starts. This matters because a broken retrieval layer will corrupt every downstream component, and it needs to be caught before users complain. Consider the cascade. The AI Data Analyst retrieves schema context, feeds it to SQL generation, executes the query, and produces a narrative. If retrieval hands over the wrong documents, every stage after that faithfully processes incorrect context into a polished but wrong answer. End-to-end metrics might still look fine: the SQL executes, the narrative is well-written. But the user gets wrong numbers. That's the silent failure that retrieval metrics exist to catch. By the end of this lesson, the right retrieval metric for a given product will be clear (precision, recall, MRR, or NDCG), along with how to compute each one, and how to connect retrieval failures to downstream quality drops. The right metric depends on what the product does with the retrieved documents, and that's a product decision, not a technical default. This lesson sits between L3.3 (signal categories) and L3.5 (semantic metrics with judges).