How the knowledge base works

THE PRINCIPLE OF RAG (RETRIEVAL-AUGMENTED GENERATION)

RAG is the technique at the heart of your knowledge base. Rather than relying only on the language model's general knowledge, your assistant looks in your entries for the passages relevant to each message written to it.

Its replies are thus grounded in what you have written, not in generalities.

SEMANTIC SEARCH

When a customer writes to you, the search starts from their latest messages: it compares the meaning of the request with the meaning of your entries, without depending on the exact words used.

For example, if a customer asks « Are you open on Saturdays? », the search can find your « Opening hours » entry even if it is worded differently.

SELECTING RELEVANT PASSAGES

For each message, the automatic search keeps at most 3 passages, and sets aside those whose meaning is too far from the request. When your assistant launches a targeted search of its own during the exchange, it can keep up to 5 passages per query.

These extracts are given to the language model, which uses them to write its reply. When a piece of information is not there, it is instructed not to invent it.

THE COMPLETE FLOW, FOR AN EMAIL

Here is the path of an email, from its arrival to the reply:

• A customer writes to you
• Their latest messages, three at most, are grouped into a single semantic search query
• A vector search is run in your establishment's search space
• The most relevant passages are kept (3 at most)
• The language model receives the message together with those extracts
• It can launch a targeted search if a detail is missing, then it writes its reply
• Depending on your setting, the reply waits for you as a draft to approve, or goes out on its own — unless a safeguard holds it back