Building a grounded assistant

1. Building a grounded assistant

You've reached into a system for data and given an agent a document to answer from. Now let's combine them in two assistants: one answers from a handbook, and the other collects details from the person asking first.

2. The pattern underneath both

Every grounded assistant does two jobs. It retrieves content that matches the question, then generates a natural answer from what it found. Validation is your job: compare a reply with the source before you trust it. Without the retrieval step, the agent would just be guessing.

3. Assistant one: Policy Lookup

Our first assistant answers the questions employees keep asking HR about travel, expenses, and remote work. We'll give it a topic, attach the policy handbook, and check that its answers come from that file.

4. Create the topic and its description

Open your agent, go to Topics, and add a topic from description with Copilot. Name it Policy Lookup and mention all three areas. Copilot tends to generalize to "policy" in the abstract, so read its example queries, add one for any missing area, and save.

5. Attach the handbook

On the Knowledge tab, click Add knowledge and upload the policy handbook. Then switch Web Search off on the Overview tab, so the handbook is the only outside source the agent can use.

6. Generate the grounded answer

You don't build the lookup yourself. With the handbook attached, the agent finds the matching passage and writes its reply from that, not from general knowledge. The order matters: retrieve first, generate second.

7. Test the Policy Lookup assistant

Open the Test panel and ask "what is our remote work policy." The reply's reference should name the policy handbook. A new file takes 5 to 10 minutes to index, so if there's no answer yet, wait and ask again. It may answer through your topic or straight from the handbook, and either is fine.

8. Assistant two: Onboarding Packet

The second assistant does a bit more. Onboarding depends on who someone's joining as, so this topic gathers a few details first, then tailors its reply.

9. Collect the inputs

Describe this topic to Copilot and name all three details: role, department, and start date. It drafts a question step for each, and each one can save its answer to a variable. Then scroll the canvas to confirm all three.

10. Tailor it to the department

The department answer decides what the reply says. Copilot often builds a separate welcome for each department, so a new hire in Marketing and one in Engineering won't hear the same thing.

11. Generate the personalized welcome

The closing reply is generated from the new hire's answers, so its wording changes every time. It might be one summary or a message per team, and either way it should reflect what they said.

12. Let's practice!

Now it's your turn to build two assistants and watch retrieval and generation work together. Almost every assistant you build will use this retrieve-then-generate pattern.

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.