Data sources are the material an agent can quote: files you upload and text you paste. Kylo splits each source into chunks, indexes them, and, on each turn, retrieves the chunks that match the customer’s message. See Add data sources for the steps.

​
The page

The list shows 10, 25, or 50 sources per page. Writes are disabled while the agent’s plan is in a non-paying state, and the page shows the billing card. Retrieval keeps working.

​
Accepted content

Each file is limited to 10 MB. The storage limit is 50 MB per agent.

​
Statuses

​
How indexing works

  1. Kylo extracts the text.
  2. It splits the text into chunks of about 500 tokens with 80 tokens of overlap, breaking on paragraphs and then on sentence boundaries.
  3. It computes an embedding for each chunk and stores it.
Re-syncing a source whose text is unchanged makes no new embedding calls. Only changed chunks are re-embedded, and a change of embedding model re-embeds everything.

​
How retrieval works

On each turn Kylo searches the agent’s chunks with the customer’s message and takes up to 6 matches, capped at 4,000 characters in total. It places them under the instructions as a knowledge block, each piece marked with its source name. The agent never sees sources that did not match. The seeded instructions tell it to use a matching piece, to say so when the answer is not there, and to name the source the way a person would.
Snippet names come from the first non-empty line of the text, cut to 80 characters. Because the agent cites sources by name, start each snippet with a clear title.

Playground

Test an answer that should come from a source.