Deterministic Output

This technique enforces a strict schema (e.g., JSON, XML, YAML) on the model's response. By providing a formal grammar or JSON Schema definition within the prompt, you eliminate ambiguity in parsing and force the model to populate predefined fields. This is foundational for API integration and data pipeline automation.

• Example: "Output your answer as a JSON object with keys: 'summary', 'key_points' (as a list), and 'confidence_score'."
• Mechanism: The model must map its internal reasoning onto the required structure, drastically reducing open-ended narrative.

This method explicitly ties all model responses to provided source material. The prompt instructs the model to derive every claim directly from the given context, prohibiting extrapolation. This is the operational implementation of the No Fabrication Rule.

• Key Instructions: "Base your answer solely on the provided document." "Do not use any prior knowledge." "For each statement, cite the relevant paragraph number."
• Use Case: Critical for Retrieval-Augmented Generation (RAG) systems, ensuring outputs are verifiable and traceable to source data.

This architecture decomposes the generation process into instructed, sequential phases. Instead of a single response, the model is prompted to generate, then verify, then correct. This introduces a deterministic fact-checking loop.

• Common Pattern:
  1. "First, draft a response to the query."
  2. "Second, review your draft. List any unsupported claims or potential inaccuracies."
  3. "Third, produce a final, corrected response."
• Benefit: Makes the model's reasoning and quality control process explicit and reproducible.

This technique uses prompt instructions to strictly define the domain, temporal scope, and verbosity of the response. It prevents off-topic elaboration and anachronisms.

• Temporal Bounding: "Only consider events that occurred before 2023."
• Domain Bounding: "Limit your analysis to financial accounting principles; do not discuss legal implications."
• Length Bounding: "Answer in exactly three bullet points."
• Mechanism: These constraints act as guardrails, reducing the model's solution space to a known, manageable region.

This prompt design controls the model's expression of certainty. It instructs the model to only state information if its internal confidence exceeds a specified level, otherwise to explicitly acknowledge uncertainty. This is a calibration prompt for honest output.

• Instruction: "If you are less than 90% confident about a fact, state 'I am not certain, but...' before providing it." "If no relevant information is in the context, say 'Cannot determine from provided sources.'"
• Outcome: Prevents the model from presenting guesses as facts, a major source of non-deterministic hallucination.

For tasks involving multiple documents, this technique provides explicit instructions for cross-referencing and conflict resolution. The prompt mandates a coherent synthesis that acknowledges and resolves discrepancies.

• Key Directives: "Compare the information in Document A and Document B." "If there is a contradiction, note it and explain which source you are prioritizing and why." "Integrate the information into a single, consistent summary."
• Benefit: Ensures deterministic output even when source materials conflict, by making the reconciliation logic an instructed, repeatable process.

A grounding prompt is an explicit instruction that requires a language model to base its response solely on provided source material, verifiable facts, or a specific knowledge base. This technique directly prevents fabrication by tethering the model's output to an authoritative reference.

• Core Mechanism: Instructs the model to act as a 'closed-book' system for the task, ignoring its internal parametric knowledge unless it aligns with the provided context.
• Example Instruction: 'Using only the information provided in the document below, answer the following question. Do not use any prior knowledge.'
• Primary Use Case: Retrieval-Augmented Generation (RAG) systems, where a retrieved document chunk serves as the mandatory source.

The no fabrication rule is an absolute, non-negotiable prohibition within a prompt. It explicitly instructs the model not to invent any details, quotes, data points, numerical values, or citations that are not present in the provided context.

• Core Mechanism: Establishes a strict boundary between permissible paraphrasing/synthesis and impermissible invention.
• Key Wording: Uses definitive language like 'Do not make up any information,' 'Only use what is provided,' or 'If the answer is not in the text, say "I cannot find that information."'
• Failure Mode: Without this rule, models often 'fill in the blanks' plausibly, which is the primary source of confident-sounding hallucinations.

A source attribution instruction mandates that a model cite the specific origin of each factual claim in its response. This forces traceability and allows for human verification, making fabrication immediately apparent.

• Core Mechanism: Requires output to include references (e.g., document IDs, page numbers, paragraph indices, URLs) inline or in a structured format.
• Example Instruction: 'For every statement of fact in your answer, cite the relevant paragraph number from the source document in brackets, like [P1].'
• Secondary Benefit: This instruction often improves model attention to the source text, as it must locate evidence for each claim before asserting it.

Structured verification is a prompt pattern that forces the model to output its internal fact-checking process in a predefined, machine-readable format. This externalizes reasoning and makes verification systematic.

• Core Mechanism: Instead of a free-text answer, the model is instructed to produce an output like a table, JSON object, or list with fields such as 'Claim,' 'Supporting Evidence,' 'Source Location,' and 'Verification Status.'
• Example Format: {"claim": "The treaty was signed in 1992.", "evidence": "...text quote...", "source": "document.pdf, page 4", "is_supported": true}
• Advantage: This decomposes the complex task of 'being correct' into simpler, auditable sub-tasks, reducing errors.

These are complementary instructions that manage a model's expression of certainty. A confidence threshold tells the model to only answer if its internal certainty exceeds a level (e.g., 'Only answer if you are >90% confident'). Uncertainty acknowledgment instructs it to explicitly state when it lacks sufficient information.

• Core Mechanism: Calibrates the model's output to avoid high-confidence hallucinations. It shifts behavior from 'guess and sound sure' to 'know or decline.'
• Example Instruction: 'If the provided documents do not contain enough information to answer the question with high confidence, begin your response with "The available information is insufficient to answer confidently."'
• Engineering Impact: Critical for production systems where a wrong, confident answer is more harmful than no answer.

This is a multi-step prompt architecture where the model is instructed to generate a response and then critique it in a subsequent step. A fact-checking loop explicitly guides the model to act as its own verifier.

• Core Mechanism: Uses prompt chaining. Step 1: 'Answer the question.' Step 2: 'Review your answer. List any factual claims. For each claim, check if it is directly supported by the source. Revise your answer to correct any unsupported claims.'
• Key Insight: Leverages the fact that a model can sometimes detect errors in its own output when the task is reframed from 'generation' to 'critique.'
• Variation: Stepwise verification breaks this into even finer-grained, instructed steps to improve reliability.