Why we ditched Vector DBs (RAG) for Perplexity's Live API

While building the AI context engine, we realized that standard RAG (Vector Databases) was too slow for dynamic content. If a user is looking at a flash sale or a live inventory count, a cached vector embedding is useless.

We decided to architect Furie using Perplexity's Sonar models to live-crawl the specific URL the visitor is on during the chat. It added about ~400ms of latency compared to OpenAI alone, but the hallucination rate on pricing questions dropped to near zero.

Curious if other AI builders here have experimented with live-browsing models vs pre-indexed vectors?