Most RAG systems retrieve what matches the words. This one retrieves what produces understanding.
The idea is simple: before retrieving anything, the system figures out what you actually need. Then it retrieves what the answer will fail without — not just what matches your keywords. Then it writes an answer that pre-answers your follow-up questions so you don't have to ask them.
Three steps of "mind-reading" before a single word of answer is written:
- Intent Reader — figures out what you actually need vs. what you literally typed
- Smart Retriever — finds what the answer will break without, ignores what looks relevant but isn't