Grounding Prompt

The most critical component is an unambiguous instruction that binds the model's output to the provided context. This mandate acts as a hard constraint, overriding the model's parametric knowledge. Effective phrasing is direct and absolute.

• Example Instruction: "Your answer must be based only on the following provided documents. Do not use any prior knowledge."
• Purpose: Establishes the source material as the sole authoritative basis for generation, creating a clear boundary that prevents extrapolation.

A grounding prompt must specify how the model should reference its sources. This transforms claims into verifiable statements and allows for easy human or automated validation.

• Defines the format: e.g., inline brackets like [Doc1, p.3], or a formal style like APA.
• Mandates granularity: Requires citations for specific facts, not just general attribution.
• Example: "Support each factual claim with a citation to the relevant document name and page number, like so: [Source_A, Section 2.1]."

Instructions must define the model's behavior when information is absent, ambiguous, or contradictory within the sources. This prevents confident fabrication.

• Uncertainty Acknowledgment: Directs the model to explicitly state "The provided documents do not contain information on X."
• Contradiction Handling: Instructs the model to identify conflicts (e.g., "Source A states Y, while Source B states Z") rather than silently choosing one.
• Confidence Threshold: Implicitly set by forbidding guesses, forcing the model to operate only on what it can directly support.

The prompt must delineate the exact boundaries of the acceptable response. This limits creative latitude and focuses the model on extraction and synthesis from the context.

• Topic Limitation: "Your response should only address the financial projections for Q3 2024."
• Format Constraint: "Output the answer as a JSON object with keys 'summary' and 'citations'."
• Length Control: "Provide a bulleted list of the top five risks, no more."
• Effect: Reduces the probability of the model 'wandering' into unsupported territory.

For complex tasks, the prompt should architect the model's internal process to separate retrieval from generation and verification. This makes the grounding logic explicit and more reliable.

• Common Pattern (Stepwise Verification):
  1. Extract: "First, list all relevant facts from the documents."
  2. Synthesize: "Then, write a summary based only on that list."
  3. Verify: "Finally, check each sentence in your summary against the extracted facts."
• Benefit: Decomposes the task, reducing cognitive load and making each step easier to audit for source fidelity.

Clear, negative instructions (prohibitions) are as important as positive directives. They close loopholes that a model might exploit.

• The No Fabrication Rule: "Do not invent any numbers, names, dates, or quotes that are not present in the sources."
• No Prior Knowledge: "Even if you know the answer, you must only use the provided text."
• No Extrapolation: "Do not draw conclusions that are not explicitly stated or logically implied by the sources."
• Function: Serves as a final safety rail, explicitly defining invalid behaviors.

For AI assisting in healthcare, grounding prompts are a safety-critical guardrail. They restrict model responses to peer-reviewed literature, clinical guidelines (e.g., UpToDate), or a specific patient's electronic health record (EHR) data. This application heavily utilizes temporal bounding (e.g., "Use guidelines published after 2020") and uncertainty acknowledgment instructions for cases where evidence is insufficient.

• Critical Function: Prevents the model from generating plausible-sounding but unverified treatment recommendations, enforcing a no fabrication rule for drug dosages or procedural steps.

Analysts use grounding prompts to generate summaries of earnings reports, SEC filings (10-K, 10-Q), or market research. The prompt mandates verbatim extraction or close paraphrasing of numerical data (revenue, EBITDA) and key statements, with source attribution to specific pages or sections. This ensures generated reports are auditable and free from speculative interpretation not present in the source.

• Common Pattern: Structured verification prompts force the model to output a table with claims, supporting quotes, and the source document page number.

News organizations and fact-checkers employ grounding prompts to verify claims against a corpus of trusted sources. The model is instructed to perform a cross-reference instruction, searching provided news archives or public databases to confirm or refute a statement. The prompt architecture often involves a fact-checking loop: first extract claims from a draft article, then ground each claim in evidence.

• Output Goal: To produce a deterministic output—a fact-check report where every verdict is tied to a specific, retrievable source.

A prompt directive that requires a model to cite the specific documents, data points, or line numbers that support each factual claim in its response. This makes the model's grounding transparent and verifiable.

• Enforces Traceability: Each claim must be linked to a source.
• Format Specification: Often includes instructions for citation style (e.g., [Doc1, Section 2]).
• Example: "For every factual statement you make, provide an inline citation to the relevant paragraph in the source material, like this: [Source A]."

An absolute, non-negotiable prohibition within a prompt that explicitly instructs the model not to invent details, quotes, data, or citations absent from the provided source material. It is the most direct guardrail against hallucination.

• Absolute Language: Uses terms like "must not," "do not invent," "strictly prohibited."
• Fallback Behavior: Instructs the model what to do if information is missing (e.g., "state 'Information not provided'")
• Example: "Do not add any information not present in the provided text. If you cannot answer from the text, respond with 'Cannot answer based on the source.'"

A prompt architecture that decomposes response generation into multiple steps, instructing the model to first draft an answer and then in a subsequent step critique and revise it for factual accuracy against the source. This separates generation from verification.

• Multi-Turn Process: Uses a chain of prompts (e.g., [DRAFT] -> [VERIFY] -> [REVISE]).
• Explicit Instructions: Guides the model to act as its own critic.
• Benefit: Catches errors the model might make in a single pass.

A prompt pattern that forces the model to output its fact-checking process in a predefined, machine-readable format. This constrains the model's reasoning to a verifiable scaffold.

• Common Formats: Tables, JSON objects, or numbered lists.
• Typical Structure: {"claim": "...", "source_paragraph": "...", "is_supported": true/false}
• Advantage: Output can be automatically parsed and validated by downstream systems.

A prompt parameter that instructs a model to only state information if its internal certainty exceeds a specified level, otherwise prompting it to express uncertainty. A calibration prompt is often used beforehand to improve the model's self-assessment accuracy.

• Threshold Instruction: "Only answer if you are >90% confident based on the source."
• Calibration Example: A prompt with known answers to tune the model's confidence reporting.
• Reduces Guesswork: Prevents the model from presenting low-probability fabrications as fact.