Bind and validate knowledge-based answers¶
After creating a knowledge base and processing its files, bind it to the target agent to test whether that assistant can use the information.
Bind the knowledge base¶
Open the target assistant’s details in Agents.
Select Add in the knowledge-base section.
Search for and select the intended library. Verify the selection marker.
Choose Done and check its name and count in agent details.
Start a new conversation from that agent.
Test a fact unique to the material¶
Avoid questions a model could answer from general knowledge. Choose a term or identifier present only in the test material. For example: “Using the bound material, give the identifier of the Bluestone guide and cite the original evidence.”
Verify the fact, document name, and supporting passage. If needed, enable developer mode and inspect the content returned by the retrieval call. See Inspect execution and troubleshoot.
Common problems¶
Problem |
Checks |
|---|---|
Library absent from the selector |
Workspace, resource existence, and active search or filters |
Library ready but answers cannot use a file |
Completed parsing and embedding, and chunks containing the target content |
Answer says only current input is allowed |
Prompt or skill restrictions on knowledge retrieval |
An old definition is cited |
Updated originals and chunks, retained conflicting documents, and semantic-rule scope |
Correct answer but uncertain retrieval |
Actual call records and sources, rather than answer plausibility |
Start a new conversation after changing bindings, skills, or material so earlier context does not obscure the result.