Two ways to reach one
They are different mechanisms, and you can use both: Attach the knowledge base to an agent. The platform retrieves relevant material and puts it in the agent’s context. The agent does not decide to search — the context is simply there. Right for an agent whose whole job is one corpus. Bind the Knowledge Retrieval capability. The agent getsknowledge_search, knowledge_get_by_id, and
knowledge_list, and calls them when it decides it needs to.
Right for a general agent that occasionally looks something up, or a gateway
exposing search to someone else’s client.
Knowledge Retrieval is scoped per binding: its configuration names which
knowledge bases that binding may search. An empty list means no access —
which is also how you expose one corpus publicly and another only internally.
Retrieval
Search is hybrid: semantic similarity and keyword matching together, not one or the other. A query about “changing my login credentials” reaches a document about password resets without needing the word to match.knowledge_search takes a query, a min_score between 0.0 and 1.0, and a
limit between 1 and 100.
Each knowledge base can carry a max results setting, capping what it
contributes per query. In a multi-knowledge-base search each contributes at most
its own cap.
Reranking
When reranking is enabled on the platform, search overfetches candidates, has a small model rescore each against the query, and returns the top results by that score. It improves precision and adds latency. Reranking is an account-level setting, not a per-knowledge-base toggle. Turn it on under Settings → Features.Supported file types
JSON is not an accepted upload type. Convert structured data to CSV or
Markdown first.