Source Attribution Prompt with Confidence Indicators

This prompt is designed for regulated and high-trust domains where every claim in an AI-generated answer must be traced back to a specific source document with a measurable confidence level. Use it when you are building a Retrieval-Augmented Generation (RAG) system for legal research, financial compliance, clinical summarization, or internal audit workflows. The prompt instructs the model to decompose its answer into discrete claims, attach a source citation to each claim, assign a confidence score, and flag any assertions that lack supporting evidence. It is not a general-purpose Q&A prompt. Do not use it when the task is creative generation, open-ended brainstorming, or when retrieved context is unavailable. This prompt assumes you have already retrieved a set of candidate documents and need to produce an auditable, verifiable output.

The ideal user is an engineering lead or AI builder integrating a RAG pipeline into a product where downstream decisions carry regulatory, financial, or clinical weight. You should have a retrieval system in place that returns document chunks with stable identifiers. The prompt expects you to supply the retrieved context, the user query, and an output schema that defines the claim structure. Before deploying, you must implement validation checks that verify every claim has a corresponding source, confidence scores fall within expected ranges, and missing-source flags are present when evidence is absent. Pair this prompt with an evaluation harness that measures attribution completeness, claim-level faithfulness, and abstention accuracy against a golden dataset of known source-answer pairs.

Avoid this prompt when your retrieval quality is unproven or when latency constraints prevent the model from processing multiple documents per request. If your use case involves creative writing, subjective analysis, or tasks where source attribution adds noise rather than trust, choose a simpler answer-generation prompt instead. For production deployments, always combine this prompt with human review checkpoints for high-risk claim categories, and log every attribution trace for auditability. The next step is to copy the prompt template, wire it into your RAG pipeline, and run it against your evaluation set to establish baseline attribution accuracy before shipping.

This template produces structured, auditable answers where every factual claim is linked to a specific source passage, assigned a confidence indicator, and accompanied by provenance metadata. It is designed for regulated domains, compliance workflows, and any system where downstream reviewers need to trace outputs back to evidence. The prompt enforces attribution completeness: if a claim cannot be sourced, the model must flag it as unsupported rather than silently generating it. Use this template as the core instruction block in your RAG pipeline, substituting the bracketed placeholders at runtime with your retrieved context, user query, output schema, and risk-level configuration.

textCopy
You are an evidence-grounded answer generator operating in a [RISK_LEVEL] domain where every factual claim must be traceable to a specific source passage. Your output will be reviewed by [AUDIENCE] who require complete attribution chains.

## INPUT

**User Query:** [QUERY]

**Retrieved Context:**
[CONTEXT]

**Output Schema:**
[OUTPUT_SCHEMA]

## CONSTRAINTS

[CONSTRAINTS]

## INSTRUCTIONS

1. Read the user query and all retrieved context passages carefully.
2. Identify every factual claim required to answer the query.
3. For each claim, locate the specific source passage that supports it. Cite the source using the format specified in the output schema.
4. Assign a confidence score to each claim using this scale:
   - HIGH: The source passage directly and unambiguously supports the claim.
   - MEDIUM: The source passage implies or partially supports the claim, but some inference is required.
   - LOW: The source passage provides weak or tangential support; the claim relies substantially on general knowledge.
   - UNSUPPORTED: No source passage supports this claim. You must still state the claim if it is necessary to answer the query, but flag it clearly.
5. If the retrieved context is insufficient to answer the query fully, generate a structured abstention response that identifies specific evidence gaps and suggests follow-up retrieval queries.
6. If multiple sources conflict on a claim, note the conflict explicitly, present both viewpoints with their respective sources, and assign confidence accordingly.
7. Do not fabricate source references. If a claim is UNSUPPORTED, set the source field to null and the confidence to UNSUPPORTED.
8. Include complete provenance metadata for every cited source: document ID, chunk index, retrieval timestamp, and a brief relevance justification.

## OUTPUT FORMAT

Return a valid JSON object matching the output schema exactly. Every claim in the answer must appear in the claims array with its corresponding attribution.

## EXAMPLES

[EXAMPLES]

Adaptation guidance: Replace [QUERY] with the user's question. Populate [CONTEXT] with your retrieved passages, each prefixed with a unique source identifier and metadata block. Define [OUTPUT_SCHEMA] as a JSON Schema object specifying the structure of the claims array, source references, and confidence enums. Use [CONSTRAINTS] to inject domain-specific rules such as citation format requirements, prohibited claim types, or mandatory disclaimers. Insert [EXAMPLES] as few-shot demonstrations showing correct attribution, conflict handling, and abstention behavior. Set [RISK_LEVEL] to high, medium, or low to adjust the strictness of the attribution requirement. Set [AUDIENCE] to describe who will review the output, which helps the model calibrate its explanation depth and formality. Before deploying, validate that the assembled prompt does not exceed your model's context window and that all source identifiers in [CONTEXT] are unique and resolvable to your document store.

The Source Attribution Prompt with Confidence Indicators is designed to sit at the end of a retrieval pipeline, receiving a set of retrieved documents and a user query, and producing a structured answer with per-claim source attribution, confidence scores, and provenance metadata. In production, this prompt is not a standalone artifact—it is a component inside a broader harness that validates the output, retries on failure, logs attribution completeness, and escalates when confidence or coverage falls below acceptable thresholds. The harness must treat the prompt's output as a structured contract, not free text, and enforce that contract before the answer reaches the user or downstream system.

Integration flow: The harness should assemble the prompt by injecting [QUERY], [RETRIEVED_DOCUMENTS] (each with a unique doc_id, source_uri, retrieval_timestamp, and content), and [CONFIDENCE_THRESHOLD] into the template. After inference, parse the output against the expected schema: an array of claim objects, each containing claim_text, source_doc_ids, confidence_score (0.0–1.0), and attribution_span. Run a schema validator that rejects malformed JSON, missing required fields, or confidence scores outside bounds. Then run a completeness validator that checks: (a) every claim has at least one source_doc_id, (b) every cited doc_id exists in the input document set, (c) no claim has a confidence score below [CONFIDENCE_THRESHOLD] without an explicit low_confidence_reason field populated. If validation fails, retry once with the error messages injected as [PREVIOUS_ERRORS] in a retry-specific prompt variant. If the retry also fails, log the failure and route to a human review queue—do not silently serve unverifiable claims.

Model choice and tool use: This prompt works best with models that have strong instruction-following and structured output capabilities (GPT-4o, Claude 3.5 Sonnet, or fine-tuned variants for regulated domains). For high-throughput pipelines, consider using tool-calling or structured output APIs (e.g., OpenAI's response_format with a JSON schema, or Anthropic's tool use with a defined ClaimAttribution tool) rather than relying solely on prompt-level format instructions. The harness should log every invocation: the assembled prompt (with PII redacted), the raw model output, the validation results, and the final served answer. This audit trail is essential for regulated domains where attribution completeness may be reviewed by compliance teams. For RAG systems with large document sets, pre-process [RETRIEVED_DOCUMENTS] to include only the top-k most relevant passages, and truncate each to a maximum character length to stay within context window limits—but preserve doc_id and source_uri metadata even when content is truncated.

Failure modes to monitor: The most common production failures are (1) claims that paraphrase source content without citing the correct doc_id, (2) confidence scores that are uniformly high despite weak evidence, and (3) missing attributions for claims the model considers 'common knowledge' but which require sourcing in regulated contexts. Instrument the harness to track attribution coverage rate (percentage of claims with at least one valid source citation) and low-confidence claim rate (percentage of claims below threshold). Set alerts when attribution coverage drops below 95% or low-confidence claims exceed 10% of a response. For high-risk domains, add a human approval gate that requires explicit sign-off on any answer containing claims below the confidence threshold or flagged with missing_source: true. The harness should never silently drop the attribution requirement—if the model cannot attribute, the system must either abstain or escalate, not fabricate citations.