Attach knowledge and choose retrieval scope

Choose retrieval scope according to the agent channel and job. Broad agent-level retrieval is convenient for voice agents with one coherent collection. Named Knowledge functions provide explicit, reusable, and source-restricted search.

Option 1: Agent-level knowledge base

For an inbound voice agent:

1. Open the agent editor.
2. Select Configuration.
3. Open Knowledge Base.
4. Select Add Knowledge Base.
5. Choose one workspace base.
6. Wait for the draft save to complete.
7. Test the draft, then publish it.

The editor shows the selected base's source count and indexed size. One base can be selected in this configuration. Removing it changes the draft but does not delete the shared knowledge base.

Wirevox compiles the base's Instructions, or its Description when no Instructions exist, into the agent's context. For Modular voice, relevant chunks are also retrieved automatically on meaningful turns and are available through the general knowledge-search tool. Realtime voice retrieves through a tool during the session.

Option 2: Reusable Knowledge function

Create a Knowledge function when you need:

• retrieval for a Website Chat Agent;
• a clearly named search capability in the prompt;
• only selected sources from a larger base;
• separately governed knowledge domains; or
• the same retrieval scope on several agents.

To configure it:

1. Open workspace Functions.
2. Create a Knowledge function with a descriptive name and purpose.
3. Select one workspace knowledge base.
4. Optionally select exact source names. No selection searches the entire base.
5. Save the reusable function.
6. Attach it from the agent's Functions panel.
7. Insert its exact [[Function: ...]] marker and write a retrieval policy.
8. Test and publish the agent draft.

The function name becomes a sanitized tool name: lowercase letters, numbers, and underscores. Insert the exact marker offered by the prompt editor rather than guessing the conversion.

Whole-base versus selected-source search

Scope	Best use	Main risk
Entire base	One coherent domain with consistent authority.	Unrelated or conflicting sources can compete in retrieval.
Selected sources	A focused product, policy, jurisdiction, language, or role.	New sources are not included automatically.
Several named functions	Distinct domains the prompt can route explicitly.	More tool choices require clearer routing rules and more tests.

Prefer source restriction when a wrong-domain answer would be costly. For example, create separately named functions for Ontario and Alberta policy sources rather than relying on the model to distinguish nearly identical documents after retrieval.

Website Chat requirement

The current Website Chat runtime assembles searchable knowledge from attached Knowledge functions. The Chat Configuration knowledge selection supplies the base's Description or Instructions to the prompt, but it does not by itself add the named retrieval tool used by chat.

Therefore, when a Chat Agent must answer from indexed content:

1. select the base in Chat Configuration if its instructions should be included;
2. create and attach a Knowledge function for actual retrieval; and
3. add the exact function marker and grounding rules to the prompt.

Test in Chat Preview and again through the published website deployment.

Source deletion behavior

If a selected source is deleted from its base, Wirevox removes that name from affected Knowledge function configurations. Review the function immediately. When its selected-source list becomes empty, its runtime behavior changes to whole-base search.

Draft, published, and shared state

Keep three lifecycles separate:

• Source state — shared indexed content in the workspace knowledge base.
• Function state — reusable search scope that can be attached to agents.
• Agent state — draft prompt, selected base, function attachments, and published snapshot.

Source edits affect the shared knowledge collection without publishing each agent. Attachment, selection, or prompt changes require saving, testing, and publishing the affected agent.

Next step

Continue with Write grounding rules.