Hybrid Catalog Search

Task

Search over a product catalog combines a keyword ranking with a vector ranking. Customers say it misses products they can describe. You have the search code, a catalog with precomputed embeddings, fifteen labeled queries and an eval that reports recall. Find what is wrong, fix one thing at a time, and be ready to defend each decision with a number.

Requirements

1. run(catalog, request) answers one request. A product is {id, sku, title, tags, category, in_stock, vec}. A query is {text, vec, filters?}; filters may set category, in_stock or both. {op: "search", query, k} returns {results}: at most k product ids, best first. Do not change the arguments.
2. Two rankings. Keyword: 5 points for each query word found in the title and 2 for each found in the tags; products with 0 points are left out. Vector: cosine similarity between the query vec and the product vec. Each ranking is best first, ties by id, and keeps its best 6. {op: "rankings", query} returns {keyword, vector} as [id, score] pairs, cosines rounded to 4 places.
3. Fusion is reciprocal rank fusion: a product gets 1 / (60 + rank) from each ranking it appears in, rank counted from 1, and the two are added. Results are ordered by that sum, ties by id, then cut to k.
4. Filters come first: a product that fails a filter is not ranked by either ranking and is never returned.
5. Recall. {op: "recall", runs} takes runs of {id, relevant, results} and returns {recall, hits, total, missed}. A run is a hit only when the relevant product id is in its results. recall is hits / total rounded to 4 places (0.0 for no runs) and missed lists the ids of the runs that were not hits. {op: "eval", queries, k} searches each labeled query {id, text, vec, filters?, relevant} and returns the same report.
6. Errors: k must be a positive integer (ValueError containing k), vec must match the catalog embedding size (vec), filters may hold only category and in_stock (filters), and any other op raises ValueError containing op.