Before you begin
A Knowledge collection organizes documents, plain text, web pages, and cloud storage folders into vector embeddings for Retrieval-Augmented Generation (RAG). Connecting a collection grounds an Agent’s responses in institutional materials, such as policy manuals, academic catalogs, or department guidelines.
Review institutional policy before uploading materials, and keep confidential records, student submissions, and unauthorized documentation out of a collection.
Steps
The interface is documented at Creating a Collection and Content Management, and connecting one to an Agent is covered under Knowledge. The decisions that matter on a campus are these:
- Name the collection for its audience and scope, not its contents. “Undergraduate advising policies, 2025-26” tells a colleague whether to reuse it. “PDFs” does not.
- Confirm a domain allows indexing before you crawl it, and set a page limit. An unbounded crawl of a campus site will pull in news archives and event listings you never wanted grounded.
- Prune the source index once processing finishes. Duplicate handbooks and superseded catalog versions are the most common cause of an Agent citing an outdated rule.
- Test with a probe question that can only be answered from one specific paragraph in your sources. A general question will appear to succeed even when retrieval failed.
Check your result
- Every uploaded source appears in the collection with a status of ready.
- A direct question in a test Chat returns the correct fact and cites its source.
- An ambiguous question does not produce invented institutional detail.
Troubleshooting
- Answers are overly broad or contradict each other. Split a large, multi-topic collection into smaller collections tailored to specific agent workflows.
- Answers are outdated or incorrect. Confirm that updated source documents were processed and that the Agent is linked to the correct collection version.
- The collection pulled in far more source material than you intended. Review document page counts, character totals, and crawled web page depth, then remove unnecessary files or restrict web crawl depth.
Next steps
To explore additional workflow guides and management patterns, visit Guides index: choose your next task.
Before you post
Before sharing prompts, configurations, or screenshots, review Community sharing and responsible AI use. Never post student records, personally identifiable information, or credentials.