Support bot (RAG + agent)

Answers customers from your own policies and price list, and does not invent what it could not find. The core of a first-line bot.

Graph preview

Вход

База знаний

RAG-поиск

LLM

Агент поддержки

Ответ

How it works

Two steps instead of one: retrieval first, then an answer grounded in what was retrieved.

Entry passes the question to Knowledge base search.
The Knowledge base node is attached to the search over a data wire — it says where to look.
Agent receives the retrieved fragments as its input message and writes the answer.
Answer returns the text.

In the Knowledge base node replace REPLACE_WITH_YOUR_KB_ID with your own base’s id (the Knowledge tab).
Tune top_k and the score threshold: too many fragments is expensive and noisy, too few makes the model answer “I don’t know” when the answer is in the documents.
Add “answer only from the retrieved text” to the agent’s prompt — otherwise the model tops the documents up from its own knowledge, and the reader cannot tell which is which.