Description
We're building a multi-tenant RAG (Retrieval-Augmented Generation) product in-house and have most of the platform up and running — ingestion, embeddings, and a working retrieval pipeline are already in place. Where we're stuck is a genuinely hard problem: guaranteeing strict, provable tenant isolation in retrieval as we scale to more customers, without tanking latency or blowing up infra cost. This is not a beginner task — we need a senior AI/ML engineer who has actually solved this class of problem before and can come in, diagnose the architecture, and fix it properly. The Problem Our current setup risks cross-tenant data leakage under certain query/retrieval conditions, and we need someone who can pinpoint exactly where and why We need an expert opinion (and implementation) on the right isolation strategy for our scale — namespace-per-tenant, metadata filtering, or separate indexes — with a clear tradeoff analysis, not just a textbook answer Retrieval accuracy and latency need to hold up even with hard isolation boundaries in place Any fix needs to be production-safe — this touches live customer data What You'll Do Audit our existing multi-tenant RAG architecture and identify isolation gaps Recommend and implement the correct isolation approach for our vector DB setup Stress-test retrieval to confirm zero cross-tenant leakage Tune for latency/cost so the fix doesn't degrade performance at scale Clearly document the root cause and the fix so our internal team can maintain it