Knowledge Bases
Give your agents something to stand on — curated knowledge bases they can search and cite.
Ground your agents.
Knowledge Bases are the documents your agents reason over — indexed for search, reachable from inside a persona’s own conversation and as an MCP tool any agent you authorize can call.
A model recalls. A citation proves.
An agent asked about your product answers from its training unless you give it something better. A knowledge base is that something: the documents you chose, searched at the moment the question is asked, so the reply carries a source instead of a guess.
What comes back, and what can’t.
An excerpt, with its source
a hit returns a short passage and the document it came from, ten per call at most, so a large corpus can’t crowd out the conversation.
Your documents, not the open web
a persona searches only the buckets granted to it, and one holding no grant at all gets nothing back rather than everything.
Update once
the bucket points at the originals instead of copying them, so one edit reaches every agent reading it.
Ingest, grant, recall.
- Ingest what you want an agent standing on; it lands as markdown rows a knowledge bucket can reference.
- Grant the persona read on that bucket. The grant names rows, not the table behind them — a document you didn’t put in the bucket stays out of reach.
- Let the agent call it — one search tool, matching titles as well as bodies. Paraphrase matching is built into the same call but stays empty until you configure an embedding service and backfill the corpus, so recall is keyword-only until you do.