Cavaridge Academy

Use chat with retrieval-grounded citations and verify the source links before relying on output.

--- title: Chat with citations status: draft note: AI-generated first-pass transcript pending video production + SME review. --- In this lesson you'll work in Studio's sandbox tenant — the **forge-msp-trio** seed gives you three sample MSP clients with realistic ticket history, licensing data, and a small docs corpus. You'll use chat, you'll get citations, and we'll talk about what to do with both. ## The flow Open Studio. Start a chat. Ask a real question — something a service-team member would ask. Try this: > Summarize the open tickets for ContosoCare and tell me which would > need a follow-up call this week. Studio responds with a summary. Look at the bottom of the response: each non-trivial claim has a citation. Click one. The cited source opens in a side panel. ## What citations actually mean A citation is a **claim of grounding**: the model is saying _I drew this from this source_. It is not a guarantee that the source actually supports the claim. Two things to check: - The cited document genuinely says what the answer says. - The cited document is **the right kind of source** for the question. A roadmap doc isn't a security finding; an old ticket isn't a current inventory. When the answer is right and the source supports it, you have a usable output. When the source doesn't support the claim, that's a hallucination even with a citation. Treat it like one. ## When there are no citations Sometimes Studio answers without citations. This is honest behavior — it's saying _I'm answering from training, not from your data_. Treat that output as a starting point, never the final answer for a customer-facing or compliance-relevant decision. If you need grounding, ask: "answer only from our connected sources, and cite each claim." ## What about retrieval-augmented chat Studio's retrieval is over the docs and data sources you've connected to the workspace. If a topic isn't connected, citations can't appear. The answer to "why didn't I get citations" is usually "we haven't connected that data source yet" — not "the AI is being lazy." ## The minimum review you should give every customer-facing output 1. Open every citation. 2. Confirm the cited source supports the specific claim. 3. Check the source is current and relevant. 4. If anything's off — even a small thing — edit before sending. This is the same standard you'd apply to a draft from a junior analyst. Studio is fast; you stay accountable. ## Hands-on exercise In the sandbox, ask three questions: - One that's well-grounded in the seed data (you should get good citations). - One that's **partially** in scope (you should see Studio decline parts). - One that's outside scope (you should see no citations or a refusal). Compare the three responses. Notice how Studio's confidence + citation pattern changes with grounding. That pattern is a tool you'll use every day.

Chat with citations

Knowledge check