Multi-Source Synthesis

This is the core instruction that directs the model to identify and reconcile discrepancies between sources. The prompt explicitly mandates a process for handling contradictions.

• Key Instruction: "When sources conflict on a point, identify the contradiction, weigh the evidence or context from each source, and synthesize the most factually consistent interpretation. Acknowledge the conflict in your final summary."
• Mechanism: The model is forced into a comparative reasoning mode, moving beyond simple extraction to evaluative synthesis.
• Example: If Document A states a project budget of $1M and Document B states $1.2M, the synthesis must note the discrepancy and may state, "Budget estimates range from $1M to $1.2M," rather than silently picking one.

A directive that requires the model to treat multiple documents as a unified corpus, checking information from one source against all others.

• Key Instruction: "For each major claim or data point, cross-reference it against all provided sources to establish consensus, corroboration, or contradiction."
• Mechanism: This prevents source isolation, where the model summarizes each document sequentially without integration. It enforces a holistic view.
• Technical Outcome: The model's attention mechanism is distributed across the entire multi-document context, building a latent representation of agreements and disagreements.

An optional but powerful mechanism where the prompt instructs the model to assign implicit authority or recency weights to different sources.

• Key Instruction: "Prioritize information from the most recent document in cases of factual conflict," or "Consider the technical specification as the primary source for product details, with the marketing brochure as secondary context."
• Mechanism: Provides a deterministic rule for conflict resolution, moving beyond simple flagging to a decision framework. This is crucial for synthesizing documents of different types (e.g., a legal contract vs. an email summary).
• Use Case: Essential in enterprise settings where a SLA document overrides a project plan, or a scientific paper overrides a news article on the same topic.

The instruction that compels the model to produce a single, logically flowing narrative from disparate parts, not a bulleted list of source summaries.

• Key Instruction: "Synthesize the information into a single, coherent narrative. Do not present your output as 'Source 1 says... Source 2 says...'"
• Mechanism: Suppresses the model's default tendency to chunk text by source. It activates planning and high-level structuring capabilities to create a new, unified document.
• Hallucination Mitigation Link: By forcing synthesis within the instruction, it reduces the risk of the model extrapolating connections that aren't present, as it must base the narrative solely on provided links.

A directive for the model to explicitly map areas of agreement and highlight missing information.

• Key Instruction: "In your synthesis, clearly identify points where all sources agree, and note topics covered by only a subset of sources."
• Mechanism: Transforms the task from pure summarization to analytical synthesis. The model must categorize information across a Venn diagram of source coverage.
• Output Format: This often produces summaries with clear sections: Unified Findings, Contested Points, and Single-Source Claims. This structured output is a direct guardrail, making the model's reasoning and source reliance transparent.

Multi-source synthesis is often combined with source attribution instructions to create verifiable outputs.

• Key Instruction: "Synthesize the information below into a summary. For each synthesized claim, cite the relevant source document using inline brackets, e.g., [Doc A]."
• Mechanism: This creates a dual-constraint prompt: the model must both create a new coherent narrative and maintain a tether to the origin of every piece of information.
• Engineering Benefit: This combination is the foundation for factual consistency checks and auditable AI outputs. It allows downstream systems or human reviewers to trace any summary statement back to its source, enabling validation and catching synthesis errors.

This application directs a model to synthesize quarterly performance data from separate departmental reports (Sales, Marketing, Engineering) into a single executive summary. The prompt must instruct the model to resolve conflicts (e.g., differing date ranges), normalize metrics, and highlight cross-departmental dependencies. For example, a prompt might specify: 'Integrate the attached Q3 reports. Where metrics conflict, use the Finance team's numbers as the source of truth and note the discrepancy in a footnote.'

Applied to logs, witness accounts, and system telemetry from a service outage or security breach to create a unified timeline and root-cause analysis. The prompt must prioritize temporal alignment and conflict resolution between automated logs and human narratives. A critical instruction is: 'Using the provided server logs, Slack transcripts, and monitoring alerts, create a minute-by-minute incident timeline. Resolve conflicts in timestamps by prioritizing the server log clock. Clearly distinguish between observed events and inferred causes.'

Consolidates information from competitor websites, press releases, financial filings, and market research reports to profile a competitive landscape. The prompt must instruct the model to separate factual claims from promotional language and to identify gaps in the available intelligence. Key instruction: 'Synthesize the provided documents about Company X's product strategy. Create a table with columns for: Claim (e.g., 'market leader in SaaS'), Source Document, and Confidence (High/Medium/Low based on source type). Note where multiple sources make the same claim.'

A prompt directive that requires a model to compare information across multiple provided sources to establish consensus and identify discrepancies before formulating a response. This is the foundational step preceding synthesis.

• Mechanism: Instructs the model to perform an explicit comparison, often outputting a preliminary analysis of agreements and conflicts.
• Example Prompt: "Before answering, compare the three provided reports on the project timeline. List all points of agreement and note any specific contradictions in dates or deliverables."

A prompt instruction that directs a model to identify and resolve conflicting statements within its own output or between its output and the provided source material. It acts as a quality check during or after synthesis.

• Key Function: Forces the model to apply logical consistency checks, a critical sub-skill for multi-source synthesis.
• Implementation: Often paired with a verification step, as in: "Generate your summary. Then, review it to detect any internal contradictions or conflicts with the source documents. List and correct them."

A prompt directive that requires a language model to cite the specific documents, data points, or references that support each factual claim in its response. This makes the synthesis process auditable.

• Output Format: Typically requires a citation format (e.g., inline doc IDs like [Doc1]).
• Purpose: Transforms a synthesized summary into a verifiable report, allowing users to trace every claim back to its origin. This is non-negotiable for legal or technical multi-source synthesis.

A prompt directive that mandates the model to support every factual assertion with specific data, quotes, or references from the provided context. This is the enforcement mechanism for source-based generation.

• Difference from Attribution: While attribution is about citing, the evidence requirement is a higher-level rule that prohibits unsupported claims. It enforces the no fabrication rule.
• Prompt Example: "For every statistic or claim in your answer, you must include a direct quote or paraphrase from one of the source documents. Do not infer or add anything not explicitly supported."

A prompt pattern that forces a model to output its fact-checking process in a predefined format, such as a table of claims and supporting evidence. This provides a transparent audit trail for multi-source synthesis.

• Common Format: A markdown table with columns for 'Claim', 'Supporting Source', 'Quote/Evidence', and 'Confidence'.
• Architectural Role: This pattern operationalizes stepwise verification, making the model's cross-referencing and synthesis logic explicit and inspectable.

An instruction that explicitly integrates or references content retrieved from an external knowledge source, grounding the model's task in that specific data. Multi-source synthesis often occurs within this broader architectural pattern.

• Relationship: In a RAG system, multi-source synthesis is the prompt-level technique used to combine the multiple retrieved documents into a coherent answer.
• System Context: The prompt instructs the model to synthesize only from the provided retrieved chunks, directly linking to contextual anchoring and bounded generation.