Bilingual Semantic Search Engine
Keyword search on a bilingual e-commerce catalog was silently missing valid matches. Replaced it with a vector pipeline on multilingual-e5-large-instruct + Qdrant, embedding Arabic and English into a single shared space. Evaluated lighter models first; e5-large-instruct won on Arabic morphology and dialect handling. Cross-language retrieval, synonym tolerance, and misspelling resilience — all resolved at the embedding layer.