Design AI Search for E-Commerce (Hybrid Retrieval, Code-Mixed)

How an e-commerce search engine combines keyword matching and dense vector search to parse Hinglish queries, rank 10 crore products, and return results in under 100 milliseconds.

Architectural blueprint for an Indian e-commerce search platform handling 6 crore daily queries. Covers Hinglish query understanding, dual-tower BM25 and HNSW vector retrieval across 16 shards, Reciprocal Rank Fusion, real-time stock freshness via outbox bitsets, and two-stage GBDT re-ranking under a 95 ms p99 latency budget.