5.3 Configuring an Agent That Has Knowledge

5.3 Configuring an Agent That Has Knowledge

Quick Answer: Configuring a high-performing declarative Copilot Agent requires four core components: (1) Scoped Knowledge, which bounds data retrieval to explicit SharePoint document libraries, OneDrive folders, specific file URLs, or public websites rather than open tenant searches; (2) System Instructions (Metaprompt), which define persona, tone, output schemas, and strict negative constraints; (3) Capability Toggles, which enable or disable Web Search (Bing), Image Generation (Designer), and Code Interpreter (Python sandbox) based on security and task requirements; and (4) Suggested Conversation Starters, which provide structured, one-click prompts that orient users and demonstrate the agent's capabilities.

While foundation Large Language Models possess remarkable linguistic fluency, their performance in specialized enterprise environments depends entirely on configuration precision. An unconfigured or vaguely guided agent risks hallucinating policies, referencing outdated organizational drafts, adopting an inappropriate tone, or drifting into out-of-scope discussions. Mastering the configuration pillars of declarative agents is a primary objective of the AB-730 certification.

---

1. Scoping Agent Knowledge: Repositories and Grounding Mechanics

In standard Copilot Chat, retrieval queries the entire Microsoft Graph accessible to the user. For a specialized agent, this broad retrieval can introduce noise—such as surfacing preliminary drafts, personal notes, or obsolete historical policies. Knowledge Scoping explicitly restricts the agent's Retrieval-Augmented Generation (RAG) pipeline to curated, authoritative data repositories.

+-------------------------------------------------------------------------+
|                    AGENT KNOWLEDGE SCOPING TIERS                        |
+-------------------------------------------------------------------------+
|  [1. SHAREPOINT SITES & LIBRARIES]                                      |
|   • Point to full site collections, subsites, or specific libraries.    |
|   • Example: https://contoso.sharepoint.com/sites/HR/Policies/          |
|   • Auto-indexes supported Office files (.docx, .pptx, .xlsx, .pdf).    |
|                                                                         |
|  [2. SPECIFIC FILE ATTACHMENTS]                                         |
|   • Pin explicit authoritative master documents.                        |
|   • Example: /Employee_Handbook_2026.pdf, /Travel_Matrix_v4.xlsx        |
|                                                                         |
|  [3. ONEDRIVE FOR BUSINESS FOLDERS]                                     |
|   • Attach curated personal or department-shared cloud folders.         |
|                                                                         |
|  [4. AUTHORIZED PUBLIC WEBSITES]                                        |
|   • Index external public domains, subpaths, or documentation hubs.     |
|   • Example: https://learn.microsoft.com/en-us/microsoft-365-copilot/   |
+-------------------------------------------------------------------------+

Supported Knowledge Sources and Ingestion Mechanics

Declarative agents support multiple distinct knowledge source types:

• SharePoint Document Libraries & Sites: Creators can attach entire SharePoint site URLs or target specific document libraries. The agent automatically leverages the Semantic Index to vectorize and index supported file formats (.docx, .pptx, .xlsx, .pdf, .txt, .rtf). When files within the document library are modified or added, the changes are automatically reflected in the agent's grounding context through Microsoft Graph synchronization.
• Specific File URLs: For razor-sharp precision, creators can specify individual file URLs. This guarantees that the agent grounds its responses exclusively on certified master assets (such as an approved corporate standard operating procedure or an annual pricing matrix).
• OneDrive for Business Folders: Useful for team-level or personal agents where project files reside in a shared cloud folder.
• Public Web Domains & URLs: Creators can ground agents in authoritative external documentation by adding specific public URLs or full web domains (e.g., regulatory compliance portals or vendor product manuals). The system crawls and indexes public text while respecting standard web robots protocols.

Knowledge Source Comparison Matrix

---

2. Realistic Business Scenario: Configuring a Global Financial Compliance Agent

Background

Woodgrove Bank requires an internal agent to assist commercial loan officers in verifying compliance with the bank's 2026 Commercial Lending Credit Policy and analyzing borrower debt-service coverage ratios (DSCR).

Configuration Architecture

1. Knowledge Scoping:
   • SharePoint Document Library: https://woodgrove.sharepoint.com/sites/CreditRisk/Policy2026/
   • Specific Master File: /Master_Loan_Underwriting_Matrix.xlsx
2. System Instructions (Metaprompt):
   • Persona: Senior Commercial Underwriting Auditor.
   • Tone: Highly formal, analytical, and risk-averse.
   • Formatting Schema: Deliver loan evaluations with (1) Executive Decision [Compliant / Non-Compliant], (2) Financial Ratios Table, (3) Risk Factor Justification citing exact policy sections.
   • Negative Constraints: "Do not approve any loan application with a DSCR below 1.25x without explicit written waiver from the Chief Risk Officer. If borrower financial data is missing from the prompt or attached workbook, do not assume baseline figures; output 'Incomplete Borrower Data' and list the missing metrics."
3. Capability Configuration:
   • Web Search: Disabled (to protect non-public financial deliberations and prevent external market chatter from influencing risk audits).
   • Image Generation: Disabled.
   • Code Interpreter: Enabled (essential for running exact Python scripts to calculate loan amortization, debt yields, and DSCR formulas).
4. Conversation Starters:
   • "Calculate the Debt-Service Coverage Ratio (DSCR) for an attached borrower income statement."
   • "Check minimum collateral requirements for commercial real estate loans over $5M."
   • "Generate a pre-underwriting compliance checklist for industrial equipment financing."

---

Exam Warning: Pay close attention to scenario questions involving numeric accuracy or data analysis in Excel. If a question asks how to guarantee mathematical precision and prevent calculation hallucinations in a declarative agent, the answer is to enable Code Interpreter.

Exam Tip: Remember that negative constraints are not suggestions; they are rigid behavioral guardrails embedded into the system metaprompt that instruct the model to refuse out-of-scope requests and avoid speculating when knowledge sources lack data.