Architect the Retrieval and Ranking Stack for Search

Query understanding first: Tokenization, spell correction, rewriting, synonym expansion, intent classification. If the query is wrong, everything after it is wrong.

Retrieval: Inverted index (BM25) as the relevance floor plus vector retrieval for semantic matching. Merge and dedupe channels.

Relevance labels: Human labels set the scale (graded relevance). Click data provides large-scale weak supervision - but clicks have position bias, which must be corrected (inverse propensity weighting; position features in a two-tower setup that are not used online).

Ranking: Explain pointwise / pairwise / listwise tradeoffs in learning-to-rank. Final ranking fuses relevance, quality, freshness, and personalization - with a hard relevance floor that personalization cannot override.