Evaluation of Context-Oriented Search Architecture Reconstruction: Offline Validation of Quality and Performance

In the segment of electronic sales of do-it-yourself (DIY) repair products, search quality significantly depends on correctly accounting for regional assortment restrictions and differences between client types. Traditional search solutions apply these restrictions at the post-search filtering stage, leading to additional computational costs, unstable response times, and inconsistencies between search, suggestions, and the catalog. This work proposes an architecture where the search query is first classified into the most probable product category, taking into account regional and user context. The obtained category is used to select a precomputed catalog index bucket $(\hat{c}, r, u)$, within which full-text search is then performed. This order shifts the assortment availability check to offline indexing, eliminating resource-intensive online filtering and ensuring predictable processing delays under high loads. The architecture's efficiency was evaluated offline by comparing users' actual purchases with the positions of the same products obtained through simulated reproduction of historical contexts. The results show that the context-oriented reconstruction of the search pipeline improves the quality of catalog ranking by nDCG@12 by 3 percentage points and simultaneously reduces service latency to around 2 ms, confirming the practical applicability of the proposed approach.