Power Platform Community Forum Thread Details

What I’m trying to achieve:

I want to deploy a public-facing Copilot agent that can answer technical questions accurately and authoritatively, using our approved technical PDFs stored in Canto document asset management system as the single source of truth.

Key constraints:

The agent must be publicly accessible (no login).
Answers must be strictly grounded in approved documents (no hallucinations).
Canto must remain the master system for document management and approvals.
The setup must scale and stay maintainable as documents change.
Document and long and contain lists and tables of data.

Proposed Approach (High level)

Because public Copilot agents cannot read SharePoint or authenticated sources, the proposal is to use Azure AI Search as the public consumption layer.

Proposed architecture:

Canto (master DAM)
   ↓ API-based sync
Azure Blob Storage (private)
   ↓
Azure AI Search (PDF extraction + vector index)
   ↓
Public Copilot Studio agent

How it works:

Approved technical PDFs are pulled from Canto via API (folders / metadata rules).

PDFs are stored privately in Azure Blob Storage.

Azure AI Search indexes and chunks the PDF content (text extraction).

The public Copilot agent queries Azure AI Search as its only knowledge source.

If information isn’t in the documents, the agent is instructed to say so explicitly.

Is this the best/right architecture for a public for this use case?

Categories:

Publish & channel management

Autonomous agents

General topics

Hi @GazzR,

yes, your proposed architecture is correct and follows Microsoft's recommended RAG pattern.

Canto → Blob Storage → Azure AI Search → Copilot Studio is the right approach. AI Search acts as the private grounding layer, the public never touches your index, and you can instruct the agent to only answer from indexed content.

A few things to check before start:

1. Tables & lists – Default chunking will break them. Add Azure AI Document Intelligence as a pre-processing step before indexing.
2. Canto sync – Build in delta sync, a deletion strategy (revoked docs must leave the index), and filter on approval state so only approved docs are queryable.
3. Citation URLs – Plan a public-friendly URL now (portal page or Canto share link). Blob storage URLs won't work for public users.
4. Hallucination guardrails – Beyond prompting, configure Copilot Studio to use knowledge sources only (disable general model knowledge).
5. Index refresh – Decide your SLA: scheduled refresh (minutes/hours) or event-driven (near real-time via Azure Function + Event Grid).

Best regards,

Valantis

✅ If this helped solve your issue, please Accept as Solution so others can find it quickly.

❤️ If it didn’t fully solve it but was still useful, please click “Yes” on “Was this reply helpful?” or leave a Like :).

🏷️ For follow-ups @Valantis.

📝 https://valantisond365.com/