AI knowledge sources, context and retrieval

Prepare website sources and uploaded materials

Choose approved web pages and documents that answer the questions your audience actually asks. Set an owner for each source, identify the current version and remove outdated or conflicting policies before making them available. Use the configured knowledge ingestion workflow for uploaded manuals and reference materials; attaching a file to an unrelated record does not add it to the agent’s knowledge.

For a website source with scheduled refresh enabled, choose a cadence that fits how often the content changes. Inspect the last successful refresh and any failed update. Content must be fetched, processed and indexed before retrieval can use it. A saved schedule, successful upload or changed source page is not proof that the updated text is searchable.

After an update, ask a question whose answer changed and inspect the retrieved material. Check that the expected document version or website text is available to the intended agent. Keep private material scoped to its audience; a public support agent should use sources approved for public answers.

Connect Knowledge Base to Agents Studio and support

Knowledge Base provides the reference material behind a useful answer. Agents Studio arranges the workflow that uses it: receive a question, retrieve relevant material, prepare an AI answer and route an unresolved request to staff. The same approved material can support customer-service guidance and selected AI agents, subject to each source’s access rules.

Explore the Knowledge Base page for the website and document workflow, and try its interactive source example. Choose Help website, Product guide.pdf or Booking policy.docx to change the question, answer and source reference. Add example source and the Daily/Weekly control illustrate source management; they do not upload a document or schedule a real website update.

Separate instructions, conversation and reference knowledge

Instructions explain the agent’s job and response rules. Conversation context supplies permitted earlier turns from the same session. Knowledge retrieval selects relevant material from configured sources. These inputs serve different purposes: adding a product manual is not the same as rewriting the agent’s instructions, and remembering a question is not permission to read another customer’s ticket.

Use approved, current reference material for facts that must remain consistent, such as setup steps or service boundaries. Give every source a clear purpose and an owner who can decide when the material needs an update.

Choose when the agent should use knowledge

The shared agent configuration supports knowledge being disabled, included automatically, searched on demand, or used in both ways. Automatic context can supply relevant material at the start of execution. On-demand access lets an enabled search tool retrieve material when the workflow requires it. The available controls depend on the agent and deployment.

A source binding can specify whether it contributes automatic context and whether it is exposed through search. Check those settings together: registering a source does not mean it is attached to every agent or used in every answer.

Select sources that match the audience

Bindings can identify a specific source or a permitted set of bindable sources. Source definitions can have their own visibility, tenant scope, classification and required permissions. Agent-private material and shared reference material should be treated according to those rules, not according to an assumption that “shared” means public on the internet.

For a website assistant, begin with material intended for visitors. For an internal agent, add operational knowledge only when the configured identity can use it. An internal note, ticket attachment or form submission does not become agent knowledge merely because it exists in the workspace.

Understand retrieval methods and citation settings

Configuration can inherit a source’s retrieval method or choose keyword, semantic or hybrid retrieval where supported. Keyword retrieval is useful for exact terminology; semantic retrieval can help match related phrasing; hybrid retrieval combines approaches. The configured search service must support the selected method.

Knowledge settings also include citation requirements and context limits. Citations make a source easier to inspect when returned by the configured workflow, but a citation requirement is not a guarantee that every answer is correct or every interface displays a link. Check the actual output and the supporting material during acceptance.

Keep reference material current

Sources can be populated from declared documents, source-owned change events or both. How you update a particular source depends on the module that owns it. Confirm which published document version or synchronized source the agent uses before concluding that an edit has reached retrieval.

Maintain a small set of questions with known answers: one common task, one exact policy detail and one question the material cannot answer. Repeat these after a substantive document or configuration change. This helps distinguish an outdated source from an unclear instruction or an access problem.

Troubleshoot a missing or unexpected answer

Inspect the selected agent version, knowledge mode, source bindings, permissions and retrieval outcome. Check whether the desired material is actually in the configured source and whether a context limit or overly broad source set is crowding out relevant content. Avoid adding unrelated documents simply to make the source larger.

When the answer requires an operation, also inspect tool permissions. Reference knowledge can explain how to change a setting; it does not itself authorize the agent to change that setting.