2.1.2. Power Platform Connectors as Knowledge
💡 First Principle: Some truth is too alive to ingest. An order's status, a ticket's owner, this morning's inventory — by the time a crawler copied them, they'd be wrong. Power Platform connectors ground the agent by querying the system of record at answer time, trading index-powered retrieval for guaranteed freshness.
This is the same connector ecosystem you know from Power Apps and Power Automate — Dataverse first among equals, plus SQL, Salesforce, ServiceNow, Zendesk, and hundreds more — repurposed as knowledge. Configure one as a knowledge source and the agent can pull the relevant records into its grounding context when a question demands them. Structured data is the sweet spot: precise fields, current values, systems whose APIs already express the queries users ask.
Two practical implications follow from the query-time design. First, connections authenticate — and how they authenticate matters, because a connection running under the end user's identity honors that user's row-level permissions, while a maker-supplied connection sees whatever the maker's account sees. Second, answer latency now includes the source system's response time, which is the price of freshness.
⚠️ Exam Trap: "Answers must reflect the current state of records in Dataverse" rules out ingestion-based sources — crawl-schedule freshness is the eliminator. Live-query grounding through a Power Platform connector is the answer shape to look for.
Reflection Question: Why would you ground "what's my order status?" through a Dataverse connector but "what's our returns policy?" through document-based knowledge — what property of each question decides it?
