I created a course Agent from a template and linked a Knowledge collection containing several lecture PDFs. I want the Agent to answer student questions using grounded material from these files and cite specific sources.
What happened
The Agent provides generic answers without referencing the lecture notes. Sometimes it states that it lacks relevant information. Other times it generates fluently written explanations using general knowledge, omitting citations entirely.
What I tried
Confirmed processing. Every uploaded PDF displays a completed processing status in the collection interface.
Started fresh. I opened a new Chat and asked the same question again.
Compared against a direct upload. I attached a single PDF to the Chat window to see whether the answer changed.
Rephrased the query to match specific headings in the source document.
This usually comes down to one of three things: a collection that is linked but not explicitly enabled on the Agent, document text that is formatted poorly for retrieval, or Agent instructions that still allow answers from general model knowledge.
Why this happens
Agents do not automatically query every file in an Organization, and retrieval-augmented generation only pulls the passages that look semantically close to the question. Both of those facts show up as the same symptom: a fluent answer with nothing from your files in it.
Common causes
The collection is linked but not enabled. A collection has to be linked, enabled, and saved on the specific target Agent.
Query wording differs from the document text. Retrieval matches on semantic similarity, so a question phrased far from the source passage may retrieve nothing useful.
Another answer path is open. Retrieval runs on a linked collection without any Safeguard, but web search, URL browsing, or instructions that permit general knowledge all give the model somewhere else to go first.
Document processing did not finish. Incomplete processing prevents passage indexing.
What to do
Enable and save the collection. Open Agent settings, verify that the Knowledge collection toggle is active, and click Save.
Check file status. On the collection details page, confirm all files show a ready status rather than processing or failed.
Probe retrieval directly. Ask a specific question that can only be answered by a unique paragraph in your document.
Name the topic in your prompt so the retrieval context has something concrete to match.
Close the other answer paths. Disable web search and URL browsing while you test grounding.
Add the Knowledge Base Constraint Safeguard if steps 1 through 5 leave the Agent still reaching past the collection. This is the case the Safeguard exists for, and it works as prompt-level guidance rather than hard technical enforcement.
Review retrieval threshold settings if they are configurable, balancing precision against broader semantic coverage.
Check a duplicated Agent’s links. If you copied an existing Agent, confirm the linked collections came across.
How to verify
Your probe question yields an answer matching exact details from the source document.
Citations appear in responses when citation settings are enabled.
Questions outside the scope of the collection are declined when Knowledge Base Constraint is active.
A new Chat session reproduces grounded answers consistently.
If it still does not work
Create a temporary collection holding a single short text snippet, which tests indexing independently of PDF formatting. Also confirm that your account has editor permissions for the Workspace, since read-only users cannot modify collection links.
If issues persist, reply with your template name, collection processing status, and web search setting. Describe the documents in general terms, keeping confidential text, student information, and credentials out of the thread.