Knowledge base synchronisation

NOTHING TO LAUNCH

When you save an entry, it is indexed straight away in our vector search engine: that is what makes it usable by your assistant. There is no button to click for this.

Each establishment has its own search space, isolated from that of other customers.

CHUNKING

A long text is cut into chunks of about 400 tokens, with an overlap of 50 tokens between two neighbouring chunks, so that the meaning is not lost at the cut.

A document that repeats the same item heading from one item to the next (« Question: », « Title: »…) may instead be cut at each of those headings: one item, one entry.

EMBEDDINGS

Each chunk is converted into a numerical vector (an embedding) by an artificial-intelligence model. That vector carries the meaning of the text: your assistant can thus find a piece of information even when the question does not use the same words as your entry.

These vectors are stored in your establishment's search space.

WHEN AN ENTRY IS NOT UP TO DATE

The « What your assistant knows » list shows the state of each entry: « Up to date », « Updating… » or « Failed ».

If indexing fails — a passing outage, for example — it is retried automatically, in the background. An entry that fails several times in a row moves to « Failed ».

As long as an entry is not up to date, a button appears above the list: « Update your assistant », with the number of entries waiting, or « Retry update ».