Lessons from deploying RAG bots for regulated industries

Built a RAG-powered AI assistant for Australian workplace compliance use cases. Deployed it across construction sites, aged care facilities, and mining operations. Here's what I learned the hard way: Query expansion matters than chunk size Everyone obsesses over chunk size (400 words? 512 tokens?). The real win was generating 4 alternative phrasings of each query via Haiku, running all 4 against ChromaDB, then merging and deduplicating results. Retrieval quality jumped noticeably - especially for domain-specific jargon where users phrase things differently than document authors.