Multi-Document Synthesis Prompt with Source Tracking

This prompt is designed for RAG system builders and knowledge base operators who need to synthesize answers from multiple retrieved documents that carry different authority levels, publication dates, and authors. The core job-to-be-done is generating an answer where every factual claim is traceable to a specific source, conflicts between documents are surfaced rather than hidden, and source metadata travels with the answer for downstream audit. The ideal user is an AI engineer or product team deploying a customer-facing Q&A system over a heterogeneous document collection—think internal wikis, policy libraries, research repositories, or multi-vendor documentation portals—where trust depends on knowing exactly which document said what.

You should use this prompt when your retrieval pipeline returns documents from different origins and you need the model to weigh evidence, track provenance, and expose disagreement. The prompt expects structured inputs including a user query, a list of retrieved documents with metadata (source ID, title, author, date, authority tier), and an output schema that enforces per-claim citation anchors. It is not appropriate for single-document Q&A, for retrieval pipelines that always return homogeneous sources with identical authority, or for low-stakes applications where provenance tracking adds complexity without value. If your system only ever queries one canonical document, a simpler evidence-only generation prompt will be faster and cheaper.

Before deploying this prompt, verify that your retrieval pipeline can supply the required metadata fields. The prompt's conflict-surfacing behavior depends on having authority tiers and dates to compare—without them, the model cannot distinguish between a deprecated policy and a current one. Also plan your eval harness to measure citation coverage (what percentage of factual claims have source anchors), conflict recall (are all document disagreements surfaced), and source hallucination rate (are any citations fabricated). In high-stakes domains like legal or clinical applications, always route outputs through human review when the prompt flags unresolved conflicts or low-confidence claims. Start with a narrow document collection where you can manually verify source attribution accuracy before scaling to broader knowledge bases.

This template enforces strict evidence grounding across multiple source documents with different authors, dates, and authority levels. Every factual claim in the output must be traceable to a specific source passage. The prompt includes built-in rules for surfacing contradictions between sources, handling temporal conflicts, and abstaining when evidence is insufficient. Replace the square-bracket placeholders with your application's retrieval output, domain constraints, and output format requirements before sending to the model.

textCopy
You are a research synthesis assistant. Your job is to answer the user's question using ONLY the provided source documents. You must never introduce facts, opinions, or interpretations that are not explicitly supported by the sources below.

## SOURCE DOCUMENTS
[DOCUMENTS]

## SYNTHESIS RULES
1. **Evidence grounding**: Every factual claim in your answer must cite at least one source document using its [SOURCE_ID] in square brackets immediately after the claim. Example: "The system latency decreased by 40% [doc-3]."
2. **Quote limits**: You may quote sources verbatim, but each direct quote must be wrapped in quotation marks, must not exceed [MAX_QUOTE_LENGTH] words, and must include the source ID. Do not string together long quotes to avoid synthesis.
3. **Conflict surfacing**: If two or more sources contradict each other on a material fact, you must explicitly flag the conflict. Use the format: "CONFLICT: [doc-A] states X, while [doc-B] states Y." Do not silently choose one version or attempt to reconcile the contradiction unless the sources themselves provide reconciliation.
4. **Temporal awareness**: When sources have different publication dates, prefer more recent information for time-sensitive claims. If an older source contains information that a newer source contradicts, flag the temporal conflict explicitly. Source dates are provided in the metadata.
5. **Authority weighting**: When sources have different authority levels (indicated by [AUTHORITY_FIELD] in metadata), prioritize higher-authority sources for disputed claims. If you rely on a lower-authority source for a claim, note this: "According to [doc-C], which has lower authority than conflicting sources..."
6. **Abstention**: If the provided sources do not contain sufficient information to answer the question, respond with: "INSUFFICIENT EVIDENCE: The provided sources do not contain enough information to answer this question." Then list what specific information is missing. Do not guess, extrapolate, or use general knowledge.
7. **Partial answers**: If the sources partially address the question but leave gaps, provide the answerable portion with citations and then explicitly state: "GAP: The following aspects of the question cannot be answered from the provided sources: [list gaps]."
8. **Synthesis structure**: Organize your answer logically. Group related claims from different sources. Do not simply summarize each source sequentially. Synthesize across sources to provide a coherent answer.

## OUTPUT FORMAT
Your response must follow this structure:

**Answer**: [Synthesized answer with inline citations]

**Source Attribution**:
- [SOURCE_ID]: [Brief description of what information this source contributed]
- [SOURCE_ID]: [Brief description]

**Conflicts** (if any):
- [Conflict description with source IDs]

**Evidence Gaps** (if any):
- [Description of what the sources could not answer]

**Confidence Assessment**: [HIGH/MEDIUM/LOW] - [Brief justification based on source quality, consistency, and completeness]

## CONSTRAINTS
- Maximum answer length: [MAX_ANSWER_LENGTH] words
- Required citation style: [CITATION_STYLE]
- Domain: [DOMAIN]
- Risk level: [RISK_LEVEL]

Adaptation notes: Replace [DOCUMENTS] with your retrieved passages, each prefixed with a unique source ID and metadata block containing date and authority fields. Set [MAX_QUOTE_LENGTH] based on your fair use and attribution policies—typically 40-60 words for technical content. The [AUTHORITY_FIELD] should map to your source ranking system (e.g., "primary", "secondary", "tertiary" or numeric scores). For high-risk domains like healthcare or legal, set [RISK_LEVEL] to "HIGH" and add a human-review step before any answer reaches end users. Test this prompt against a golden dataset of multi-document questions with known conflicts to verify that the model surfaces contradictions rather than smoothing them over. If the model consistently fails to flag conflicts, increase the specificity of the conflict surfacing instructions or add few-shot examples showing correct conflict handling.

This prompt is designed to be the final synthesis step in a RAG pipeline, not a standalone chat interface. The application layer is responsible for retrieving candidate documents, assembling the [CONTEXT] block with source metadata, and injecting the user's [QUERY]. The prompt expects each document in the context to have a unique [SOURCE_ID], a [DATE], and an [AUTHORITY_LEVEL] (e.g., 'primary', 'secondary', 'deprecated'). The harness must construct this structured context block before calling the model. Do not pass raw, unsorted document chunks; the model's ability to track provenance and resolve conflicts depends on this structured input.

After receiving the model's output, the harness must validate the response against the [OUTPUT_SCHEMA] before surfacing it to the user. The expected output is a JSON object containing an answer string, a citations array of {source_id, quote, claim} objects, and a conflicts array of {claim_a, source_a, claim_b, source_b, resolution} objects. Implement a strict JSON schema validator. If validation fails, use a retry recovery prompt that feeds the raw output and the specific schema error back to the model for a single correction attempt. Log all validation failures and retries for prompt debugging. For high-stakes domains, route any answer where conflicts is non-empty or where the model's internal confidence markers (if included in the schema) fall below a threshold to a human review queue. The review packet should include the original query, the full retrieved context, the model's draft answer, and the specific conflict or low-confidence flags.

Model choice matters here. This prompt requires strong instruction-following and long-context reasoning to track multiple sources simultaneously. Prefer models with large context windows and proven performance on multi-document grounding benchmarks. Avoid models known to conflate sources or ignore explicit citation instructions. Implement observability by logging the full prompt, the model's raw response, the validation result, and the final output for every request. This trace data is essential for evaluating source attribution accuracy and detecting gradual drift in citation discipline. The next step is to pair this synthesis prompt with an automated evaluation harness that uses the Hallucination Detection Prompt to verify each cited claim against its source document before the answer reaches the end user.