Knowledge Base
Unlike an ordinary Agent that "retrieves once per turn on a fixed schedule," LoopAgent turns knowledge retrieval into a tool the model decides on itself: whether to search, how many times, and with what keywords are all judged by the model during the conversation.
Two retrieval tools
Once you bind at least one knowledge base to a LoopAgent, the model automatically gains two capabilities:
| Capability | Effect |
|---|---|
| Autonomous retrieval | The model organizes its own query terms to search the knowledge base; when results are unsatisfactory, it automatically refines the wording, changes terms, or lowers the threshold and searches again. |
| Deep-read documents | After a hit, it can read a whole document deeply, or open a window by keyword within a long text to locate content — something fixed retrieval cannot do. |
These two tools appear only when a knowledge base is bound. Without a bound knowledge base, the model does not retrieve.
Retrieval scope and parameters
- Knowledge-base scope: the model can only search within the allowlist of knowledge bases you configured; going out of bounds is blocked.
- Recall count: the model can adjust it within a range, but the number of items that ultimately enter the answer is capped.
- Relevance threshold: high by default to guarantee quality; on a zero-hit, the model can lower the threshold on its own and retry.
- Search method, knowledge graph, reranking, and so on are decided by you in the configuration and cannot be changed by the model.
A zero-hit is not an error: when nothing is found, it returns normally and prompts the model to change wording, lower the threshold, or switch bases and retry.
Knowledge recall and inline citation markers
- Knowledge recall: requires "Show sources" to be turned on in the configuration. LoopAgent accumulates and shows recalled sources for every retrieval (an ordinary Agent shows only the last one); this is not duplicate rendering.
- Inline citation markers (such as
[1]): require both conditions — "Show sources" turned on, and the source display style set to "labels." With any other style, there is only a recall list and no inline markers.
Tip
When knowledge is insufficient, you can emphasize in the Persona that the model should "deep-read the full text of a retrieved document when necessary," which raises the trigger rate of the deep-read tool and makes answers more thorough.