Multilingual RAG System
A production retrieval system that answers in Arabic and English for government and enterprise clients, built around the parts of Arabic that generic pipelines get wrong.
- 5+
- Enterprise clients
- <1.5s
- P95 latency
- 2
- Languages served
What had to change
Retrieval stacks are built for English. Arabic breaks them in specific ways: right-to-left text, diacritics that change meaning, dialect variation across the Gulf, and chunking that has to respect a script whose word boundaries do not behave like Latin ones. Regulated clients also need every answer traceable back to a source.
What Louay built
Louay architected the pipeline end to end — bilingual chunking, RTL-aware processing, a multi-layer cache in front of retrieval, guardrails for regulated environments, and tracing so each answer can be inspected after the fact.
What the work enabled
A system serving both languages from one pipeline, with the caching and retrieval layers designed for low-latency answers under regulated conditions.