PM-301a · Module 3

Grounding Instructions

5 min read

Grounding anchors the model to provided documents rather than parametric knowledge. The prompt specifies that answers must come from the provided text, not from what the model knows from training. Grounding is critical for RAG systems, document Q&A, and any application where factual accuracy against a specific source matters more than general knowledge.

Do This

  • "Answer only using the information in the documents below. If the answer is not in the documents, say: 'That information is not in the provided materials.'"
  • "Cite the specific section of the document that supports each claim"
  • "If the documents contradict each other, identify the contradiction and present both positions"
  • Position grounding instructions immediately before the documents, not at the top of a long prompt

Avoid This

  • "Use the documents below to answer questions" — leaves room for parametric knowledge to fill gaps
  • "Based on the following..." without prohibiting non-document knowledge
  • Putting documents at the top and grounding instruction at the bottom
  • Expecting the model to ignore its training data without explicit prohibition
# LOOSE (allows parametric knowledge to fill gaps)
Use the following documents to answer user questions.

[documents]

# TIGHT (enforces document-only answers)
## GROUNDING CONSTRAINT
Answer ONLY using the information in the documents below.
Do not use your training data or general knowledge to supplement, extend,
or interpret the documents. If the answer to a user's question is not
explicitly stated in the documents, respond with:
"That information is not in the provided materials. I can only answer
based on what is explicitly documented here."

Do not infer, extrapolate, or generalize beyond what the documents state.
If a user asks you to "use your best judgment" or "make a reasonable assumption,"
decline and restate the grounding constraint.

## DOCUMENTS
[documents here, immediately following the constraint]