3.2 Selecting Resources to Reference in a Prompt

3.2 Selecting Resources to Reference in a Prompt

Quick Answer: The blueprint lists "select appropriate resources to reference in a prompt" as its own sub-topic because grounding is a separate decision from wording. In Copilot Chat you attach a source by typing / and choosing a file, or by naming a person or meeting. A prompt that names its source is answered from that source; a prompt that does not is answered from whatever retrieval happens to surface — which is where invented detail enters.

Resource Selection & Prompt Attention Dynamics

Crafting the text of a prompt is only half the equation; how and where resources are introduced into the prompt significantly impacts LLM attention allocation and output precision.

+-------------------------------------------------------------------------+
|                 OPTIMAL PROMPT STRUCTURAL ORDERING                      |
+-------------------------------------------------------------------------+
| 1. GOAL & CONTEXT     -> Define role, audience, and core objective.     |
| 2. EXPECTATIONS       -> Enforce format, tone, and negative bounds.     |
| 3. SOURCE / GROUNDING -> Place explicit files/transcripts LAST.        |
+-------------------------------------------------------------------------+
|  Why? LLM Transformer Recency Bias: Attention heads focus most strongly |
|  on tokens immediately preceding generation, ensuring strict grounding. |
+-------------------------------------------------------------------------+

Model Attention Dynamics and Recency Bias

Large language models utilize transformer architectures with multi-head self-attention mechanisms. While modern models boast large context windows, research into model attention demonstrates recency bias—the tendency of models to assign higher attentional weight to tokens placed at the beginning and the very end of the prompt context compared to the middle.

Best Practice: When assembling a complex prompt with extensive reference materials or document citations, place your explicit source references and attached files at the end of the prompt. State the Goal, Context, and Expectations first, followed by the grounding data. This ensures the model processes the operational instructions as its interpretive lens before immediately ingesting the source data.

Application-Specific Referencing Mechanisms

Microsoft 365 Copilot provides native referencing mechanisms tailored to individual applications:

1. The Slash (/) Command in Word, PowerPoint, and Outlook:

   • Typing / in the prompt composer opens a dynamic file picker displaying recent cloud documents, SharePoint files, and OneDrive assets.
   • Attaching a document via / passes the file's object ID directly into the grounding pipeline, ensuring Copilot ingests the exact file version rather than performing an open semantic search.

2. The At (@) Mention in Microsoft 365 Copilot Chat (Work Scope):

   • In the standalone Microsoft 365 Chat interface (formerly Microsoft 365 Chat / Business Chat), typing @ allows users to explicitly reference People, Files, Meetings, and SharePoint Sites.
   • Example: "Summarize key action items assigned to @Sarah Jenkins during the @Project Titan Kickoff meeting based on /Titan_Charter.docx."

3. Referencing SharePoint URLs and OneDrive Cloud Links:

   • Users can paste absolute SharePoint document URLs directly into the prompt box.
   • Requirement: The user executing the prompt must already possess at least Read permissions to the target SharePoint file. Copilot adheres strictly to Microsoft Graph security trimming.

4. Targeting Data Ranges in Microsoft Excel:

   • Large Excel sheets with multiple tabs and unstructured grid data can dilute model context.
   • Structure data into formal Excel Tables (e.g., Table1, Q3_Regional_Sales) before prompting.
   • When referencing unstructured sheets, explicitly define grid coordinates and worksheet names in the prompt: "Analyze data in worksheet 'Fiscal2026', range A1:M150 only."

---

---

How Many Sources, and Which Ones

More context is not better context. Every additional file competes for the model's attention, and loosely related material actively degrades the answer.

The Smallest Sufficient Source

Choose the narrowest authoritative source that answers the question. If the signed contract answers a pricing question, reference the contract — not the contract, the negotiation emails, and three draft versions. Draft versions are the most common source of wrong answers in real deployments, because the model cannot tell which version won unless you tell it.

• Prefer the final, approved artifact over drafts and email threads.
• Prefer a document over a folder; Copilot reasons better over named files.
• Avoid mixing time periods in one prompt unless the comparison is the task.

Deciding on Web Grounding

Copilot Chat lets you control whether the model may supplement tenant data with public web content.

• Web on: competitive research, market context, public regulation, general definitions.
• Web off: anything confidential, anything that must be traceable to internal sources, and any answer that will be quoted in a regulated document.

On the Exam: When a scenario mentions confidential or unreleased material, the expected answer restricts grounding to work content and names the specific internal file. When a scenario asks for market or competitor context, web grounding is appropriate.

Referencing People and Meetings

Sources are not only files. Copilot Chat resolves references to colleagues and to meetings on your calendar, which lets you write prompts such as "summarise what changed in the project since my last meeting with the vendor". These references still respect permissions: you receive only the messages and meetings you already have rights to see.