01 / RETRIEVALAn answer with a source.
Your documents are split into passages and each one is turned into a vector. That is what lets it match on meaning instead of keywords. A question retrieves the closest passages, a second pass re-ranks them, and only the best few reach the model.
Every answer keeps a pointer back to the passage it came from. When something reads wrong, you fix the paragraph that caused it, not the prompt.
02 / MODEL ROUTINGThe right path for the question.
Not every message needs the same model. Classifying intent is cheap work and runs on a small one; the reply that a customer reads runs on a strong one.
The chain has a fallback. If a provider slows down or goes dark, the next one picks up mid-conversation instead of leaving someone staring at a spinner.
03 / TOOLSReal actions. Real records.
Anything factual is a function call, not a guess: real calendar availability, a live order status, a write into the CRM. Each tool has a typed schema, so the model cannot invent a parameter.
Writes are made repeatable. If the same call fires twice, you get one record, not two.
04 / CHANNELSOne person. One conversation.
The website widget, WhatsApp Business and Instagram all speak to the same brain through their own APIs. Each has its own rules about templates, session windows and who may message first.
Identity is tied to the person, not the channel. Someone who asks on the site at night and follows up on WhatsApp is one thread, not two.
05 / GUARDRAILS AND EVALSA person when it matters.
Below a confidence threshold it stops and offers a person. Refusing is a designed outcome, not a bug: it is cheaper than a wrong price quoted in writing.
Before any change ships, it runs against a set of real questions with known answers. Personal data is masked on the way in and retention is set to a number you choose.