A user searches for "refund policy” but the doc they need is titled "returns and reimbursements." And keyword matching scores it near zero. Vector embeddings fix that by representing meaning as numbers, so similar meanings land near each other in space. 🧵