Source-Based Generation

This is the foundational, absolute prohibition. The prompt explicitly instructs the model not to invent any details—including names, dates, statistics, quotes, or citations—that are not present in the provided source material. It transforms the model's role from a generative storyteller to a precise extractive summarizer or paraphraser. For example: "Your response must contain ONLY information present in the provided documents. Do not add any details, examples, or conclusions not explicitly stated."

The prompt mandates that every factual assertion be supported by specific evidence from the source. This moves beyond general grounding to traceable attribution. Common instructions include:

• "For each key point, cite the relevant paragraph number or document section."
• "Support all claims with direct quotes or paraphrases from the text."
• "If the source does not contain information to answer the query, state 'Information not found in provided sources.'" This creates a self-documenting output where the provenance of every claim is clear.

The prompt strictly limits the domain of the response to the content and context of the provided sources. This involves:

• Temporal Bounding: Confining answers to events/data within the timeframe covered by the sources (e.g., "based on the 2023 annual report").
• Topical Bounding: Directing the model to ignore related but external knowledge (e.g., "Do not apply general knowledge; use only the provided clinical trial data.").
• Role Bounding: Casting the model as a neutral reporter of the source, not an analyst adding external interpretation. This reduces extrapolation and unsupported inference.

To make verification machine-readable and consistent, prompts enforce deterministic output formats. This reduces creative latitude and ensures the model separates claims from evidence. Examples include:

• Tabular Output: "Present your answer as a table with columns: 'Claim', 'Supporting Quote', 'Source Document'."
• Structured JSON: "Output a JSON object with keys 'summary', 'supporting_evidence', and 'citations'."
• Stepwise Protocols: Instructions like "First, list all factual claims. Second, for each claim, provide the supporting text." This stepwise verification architecture makes the fact-checking process explicit and auditable.

Sources can be ambiguous or contradictory. Effective source-based generation prompts include procedures for handling these cases:

• Contradiction Detection: "If sources present conflicting information, note the conflict and describe the differing statements."
• Multi-Source Synthesis: "Integrate information from all provided documents. Where they agree, state the consensus. Where they differ, note the discrepancies."
• Uncertainty Acknowledgment: "If the sources are incomplete or unclear on a point required by the query, explicitly state 'The provided sources do not offer sufficient information to confirm this.'" This prevents the model from filling gaps with plausible guesses.

In legal tech, source-based generation is used to automate the analysis of case law, contracts, and regulatory documents. The model is instructed to base all summaries, risk assessments, and clause comparisons exclusively on the provided texts.

• Key Instruction: 'For each claim about liability or obligation, cite the specific clause and document section.'
• Example: A system ingests multiple merger agreements and is prompted to 'List all non-compete clauses, paraphrasing directly from the source documents. Do not infer any terms not explicitly stated.'
• Outcome: Produces a deterministic table of clauses with direct citations, enabling lawyers to verify every statement.

Healthcare applications use this methodology to generate structured patient summaries or insurance prior authorization letters from unstructured doctor's notes and lab results.

• Key Instruction: 'Populate the SOAP (Subjective, Objective, Assessment, Plan) note template using only the data points from the provided patient chart. If a required field has no corresponding data, output "Not documented."'
• Example: A model is given a day's clinical notes and prompted to create a discharge summary. It must derive medication lists, vital signs, and treatment plans verbatim or via strict paraphrase from the notes.
• Outcome: Ensures no medical details are invented, which is critical for patient safety and regulatory compliance (e.g., HIPAA).

Investment firms deploy source-based generation to analyze quarterly reports, SEC filings, and earnings call transcripts. The model is constrained to extract and synthesize figures and statements without extrapolation.

• Key Instruction: 'Calculate the year-over-year revenue growth mentioned in the provided CEO statement. Use only the numbers explicitly stated in the transcript. Show your calculation.'
• Example: Given multiple analyst reports and a 10-K filing, the prompt is: 'Synthesize the three primary risk factors listed in the Management's Discussion & Analysis section. Do not introduce risks from other sources or general knowledge.'
• Outcome: Creates auditable summaries where every financial metric is traceable to a source line, preventing misrepresentation.

Enterprise help desks implement source-based generation to answer employee IT questions using only the latest internal documentation, policy PDFs, and software manuals.

• Key Instruction: 'Answer the user's question about VPN configuration using only the 2024 IT Policy Guide and the 'Zscaler Admin Manual' provided. If the answer is not found, state: "This is not covered in the provided documentation."'
• Example: A user asks, 'How do I request access to the CRM?' The model searches provided HR and IT guides to output the exact steps, required form numbers, and approval chain.
• Outcome: Drastically reduces incorrect or outdated instructions, as the model cannot rely on its potentially obsolete training data.

Researchers use this technique to draft literature review sections or compare findings across a provided corpus of academic papers.

• Key Instruction: 'Compare the methodologies used in Paper A and Paper B for measuring protein binding affinity. Describe each method in your own words, but ensure every descriptive element is directly supported by a quote from the respective papers' Methods sections.'
• Example: A model is given 20 PDFs on climate models and prompted: 'List the three most cited limitations of current models, as stated in the provided papers' Conclusion sections. For each limitation, cite the paper(s) that mention it.'
• Outcome: Produces a draft with high citation integrity, where every claim is anchored to a source, ready for researcher verification.

Media organizations employ source-based generation to create fact-checking reports or unbiased summaries of events by grounding the model in primary source materials like press releases, official statements, and raw data.

• Key Instruction: 'Based on the provided press release from Company X and the regulatory filing from the SEC, generate a timeline of events regarding the product recall. Only include dates and details that appear in both or are uncontradicted.'
• Example: Given multiple news articles on a developing story, the prompt is: 'Identify the claims that are consistently reported across all three provided articles. Ignore claims that appear in only one source.'
• Outcome: Generates a consolidated report that highlights verified, multi-sourced information and flags discrepancies, reducing the spread of single-source misinformation.

A grounding prompt is the foundational instruction that explicitly requires a language model to base its response on provided source material, verifiable facts, or a specific knowledge base. It is the primary directive that initiates source-based generation, preventing the model from relying on its parametric memory, which may be incomplete or outdated.

• Core Function: Acts as the initial constraint, tethering all subsequent model reasoning to the provided context.
• Example Instruction: "Answer the following question using only the information provided in the document below. Do not use any prior knowledge."

An evidence requirement is a prompt directive that mandates the model to support every factual assertion with specific data, quotes, or references extracted directly from the provided source texts. It operationalizes the grounding prompt by forcing traceability.

• Key Mechanism: Transforms a general grounding rule into an actionable, step-by-step process for the model.
• Output Format: Often results in responses that interleave claims with inline citations (e.g., "According to the report [Doc1, p.4], quarterly revenue grew by 15%.").

The no fabrication rule is 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 context. It is the strictest form of a hallucination guardrail.

• Zero-Tolerance Policy: Often phrased as "If the information is not in the source, do not generate it under any circumstances."
• Critical for Compliance: Essential in legal, medical, and financial applications where unsupported speculation constitutes a material risk.

A retrieval-augmented prompt is an instruction that explicitly integrates or references content retrieved in real-time from an external knowledge source (e.g., a vector database or search API). It is the architectural precursor that provides the 'source' for source-based generation.

• System Architecture: Combines a retrieval system (to find relevant documents) with a precise grounding prompt (to generate from them).
• Dynamic Context: Enables source-based generation on live, proprietary, or updated data beyond the model's training cutoff.

A self-verification prompt is an instruction that guides a model to act as its own critic, systematically checking its initial draft response for errors, inconsistencies, or unsupported claims against the source material. It adds a recursive quality check to the generation process.

• Process: Often implemented as a multi-turn or chain-of-thought prompt: "First, write an answer. Second, review each sentence and cite its source. Third, revise any unsupported statements."
• Enhances Fidelity: Catches subtle hallucinations that may slip through a single-pass generation.

Structured verification is a prompt pattern that forces a model to output its fact-checking process in a predefined, machine-readable format. This makes the model's adherence to source-based generation explicit and auditable.

• Common Formats: Requires outputs like tables with columns for Claim, Supporting Source Text, and Page/Line Reference.
• Engineering Benefit: The structured output can be automatically parsed and validated by downstream systems, providing a deterministic check on the generation process.