This exercise demonstrates the lost-in-the-middle effect and shows how grounding fixes it. It also trains participants to think like system designers.
Step 1: Create a long-context task
-
Upload a 30-page document to a workspace.
-
Ensure the document is readable and contains structured headings.
Step 2: Hide a needle in the middle
-
Insert a specific, unusual fact in the middle of page 15.
-
Example: “The project code name is Blue Giraffe.”
-
Ensure it appears only once.
Step 3: Test memory under long context
-
Ask complex questions about page 1 and page 30.
-
Confirm the agent answers correctly.
-
Then ask: “What is the project code name?”
Step 4: Apply a grounding fix
If the agent fails, repeat the task using a grounding prompt such as:
- “Focus specifically on page 15 of the attached artifact to find the project code name. Quote the exact line where it appears.”
Step 5: Record observations
Participants should record:
-
whether the model found the fact without grounding
-
whether the model found the fact with grounding
-
what changed in the prompt structure
-
whether retrieval improved accuracy