Knowledge Retrieval
Two ways your agent can use attached collections — automatic context or an on-demand search tool.
Agents with attached collections can use that knowledge in two ways. You pick the mode per agent in the builder.
Automatic (default)
On every user turn, the platform searches the attached collections semantically and quietly injects the most relevant snippets into the agent's context. The agent "just knows" things — no visible searching. This is the right default for agents whose every answer should be grounded in your documents.
Search tool (on demand)
The agent gets a search_knowledge_base tool and decides for itself when to
use it — like a person reaching for the manual only when a question calls for
it. Searches are keyword-based (names, product codes, error messages match
especially well) and skip the embedding step entirely — no per-turn cost when
the agent doesn't need the knowledge base.
Good fit: agents that chat casually most of the time and only occasionally need to look something up.
Worth knowing:
- One mode per agent. Switching to the search tool turns automatic injection off — the agent won't receive both.
- Form steps don't search. In search-tool mode, knowledge is available in conversation steps only. If a form step in your flow needs document answers, keep the agent on Automatic.
- Results always come from the same documents you uploaded to the attached collections; access is limited to the collections attached to the agent.
Switching modes
- Open your agent in the visual Agent Builder.
- In the settings drawer, open the Runtime tab and expand the Knowledge Retrieval section. (The separate Knowledge tab is where you attach collections — both matter: attach there, pick the mode here.)
- Pick Automatic (each turn) or Search tool (on demand) and save.
- Publish the agent — the mode applies from the next session.
Keyword search in workflows
The same keyword engine is available as a workflow action: Search Collection (Keyword). It takes collections, a query and a limit, and returns matching chunks — like the semantic Search Collection action, but keyword-based and without embedding cost. Scores are relevance weights (higher = better), not 0–1 similarities.