Kata CI GuideKnowledge
Ground agents in your organization's trusted documents.
Knowledge is what keeps agents honest. Instead of answering from a model's general training, a grounded agent retrieves from documents you chose, and cites them.
Knowledge bases
A knowledge base is a named collection of ingested documents, owned by
the workspace and bound to any number of agents. Organize by domain, not by
agent: procurement-policies, product-catalog, hr-procedures. One
domain, one knowledge base, reused everywhere.
Manage them under Console → Resources → Knowledge.
Ingesting documents
Upload sources: PDF, Word, Excel, Markdown, plain text. Scanned documents go through text recognition automatically.
Wait for ingestion. Each source moves through processing to ready; you can watch status and chunk counts per source. Failures are shown per document, with the reason.
Spot-check with real questions. Ask a bound agent something specific from a document you just added and confirm the citation points where you expect.
The Console home dashboard tracks knowledge health for the whole workspace: chunk counts, ready rates, and last-ingested times per knowledge base, so stale or failed ingestion is visible without opening each base.
How retrieval works
When a user asks something, the agent searches bound knowledge bases with a hybrid of semantic and keyword matching, reranks the candidates, and reasons over the best passages. Answers cite their sources so users can verify.
Per knowledge base you can tune retrieval settings, such as how many passages are retrieved and how documents are chunked. Defaults are sensible; change them only in response to observed retrieval issues.
Keeping knowledge current
Knowledge is a living asset:
- Re-upload a document to replace its old version, and remove sources that no longer apply. Stale policy is worse than no policy.
- When users report wrong answers about facts, check knowledge first: is the right document there, did it ingest cleanly, does retrieval find it?
- Review coverage periodically against what users actually ask. Traces show which questions found nothing.