Building a grounded assistant

Building Agents with Microsoft Copilot Studio

Nia Joseph

AI Product Manager, Microsoft

The pattern underneath both

 

 

  • Retrieve → find matching content in an approved source
  • Generate → write a natural answer from it
  • Validate → check the reply matches the source

 

Grounded assistants retrieve then generate, with validation

Building Agents with Microsoft Copilot Studio

Assistant one: Policy Lookup

$$

Policy Lookup: question to agent to source to grounded answer

 

 

  • Answers HR policy questions
  • Travel, expenses, and remote work
  • Question → policy handbook → grounded answer
Building Agents with Microsoft Copilot Studio

Create the topic and its description

 

 

  • Topics → Add a topic → from description with Copilot
  • Describe several policy areas
  • Add missing examples → Save

 

 

Add a topic from a description, with Copilot drafting it

Building Agents with Microsoft Copilot Studio

Attach the handbook

 

Copilot Studio agent Knowledge tab with the Add knowledge button

 

 

  • Knowledge tab → Add knowledge
  • Upload the policy handbook
  • Web Search off → the file is the only source
Building Agents with Microsoft Copilot Studio

Generate the grounded answer

 

 

  • The agent looks up the handbook
  • Then writes from retrieved content, not general knowledge
  • Order matters: retrieve first, generate second

 

Retrieve first, then generate from retrieved content

Building Agents with Microsoft Copilot Studio

Test the Policy Lookup assistant

 

A grounded answer about the remote work policy, with a reference to the policy handbook

 

 

  • Open the Test panel
  • Ask "what is our remote work policy"
  • Reference names the handbook
  • No answer yet? Indexing takes 5 to 10 minutes
Building Agents with Microsoft Copilot Studio

Assistant two: Onboarding Packet

 

 

  • Does a bit more than the first
  • Gathers a few details first
  • Same generate step, with a collection step out front

Collect details, retrieve, generate an onboarding packet

Building Agents with Microsoft Copilot Studio

Collect the inputs

$$

Three Question nodes for role, department and start date, each saving its answer to a variable

 

 

  • Three question steps: role, department, start date
  • Each question step can save its answer to a variable
  • Collect all three before the summary
Building Agents with Microsoft Copilot Studio

Tailor it to the department

 

 

  • The department answer decides the reply
  • Often a separate welcome per department
  • Marketing's welcome ≠ Engineering's

The department variable selects the right checklist

Building Agents with Microsoft Copilot Studio

Generate the personalized welcome

 

Role, start date, and department items converge into a personalized summary

 

 

  • The closing reply is generated
  • Built from the new hire's answers
  • Wording varies from run to run
Building Agents with Microsoft Copilot Studio

Let's practice!

Building Agents with Microsoft Copilot Studio

Preparing Video For Download...