Legal RAG Answer with Citation Prompt Template

This prompt template is designed for legal AI builders who need to generate answers strictly grounded in retrieved case law, statutes, or contract clauses. It is intended for production RAG systems where every claim must be traceable to a source, jurisdiction must be respected, and uncertainty must be surfaced explicitly. The ideal user is an engineering lead or AI architect integrating a legal research or contract analysis feature into a product used by legal professionals. The core job-to-be-done is transforming a set of retrieved legal documents and a user question into a verified, citation-backed answer that a lawyer can audit, not a summary that merely sounds plausible.

Use this prompt when your system reliably retrieves relevant legal documents and you need an answer that includes direct quotes, jurisdiction-aware citations, and explicit flags for missing or conflicting authority. It is appropriate for workflows like legal research memo drafting, contract obligation extraction, and compliance Q&A over regulatory text. The prompt assumes that the retrieval step has already occurred and that the provided context contains the universe of permissible evidence for the answer. It enforces a strict contract: the model must not introduce external legal knowledge, must cite to specific sections or page numbers, and must separate binding authority from persuasive commentary.

Do not use this prompt for general legal advice, consumer-facing chatbots without human review, or workflows where the retrieval step is unreliable. If your system cannot guarantee that the provided context contains the answer, this prompt's abstention and uncertainty mechanisms will trigger frequently, which may degrade user experience. This prompt is also unsuitable for e-discovery review, privilege log generation, or any workflow requiring subjective legal judgment. In high-stakes deployments, always pair this prompt with a human review step and automated eval checks for citation hallucination, misattributed holdings, and jurisdictional mismatch before any answer reaches an end user.

This prompt template is designed for legal AI builders who need to generate answers strictly grounded in retrieved case law, statutes, regulations, or contract clauses. It enforces citation discipline by requiring every factual claim to be paired with a source reference, a direct quote, and a jurisdiction note. The template uses square-bracket placeholders that your application must populate before sending the request to the model. Replace each placeholder with your system's live data: the user's question, the retrieved context chunks, your desired output schema, any domain-specific constraints, and the risk level that controls how the model handles uncertainty.

textCopy
You are a legal research assistant. Your answers must be grounded exclusively in the provided [CONTEXT] and must never rely on your own knowledge of law, cases, or statutes.

## INPUT
[INPUT]

## CONTEXT
[CONTEXT]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "answer": "string (the synthesized answer)",
  "citations": [
    {
      "source_id": "string (from context metadata)",
      "source_name": "string (case name, statute title, contract section)",
      "jurisdiction": "string (court, state, federal circuit, or regulatory body)",
      "direct_quote": "string (verbatim text from the source)",
      "relevance": "string (one sentence explaining why this source supports the answer)"
    }
  ],
  "uncertainty_flags": [
    {
      "type": "string (one of: 'conflicting_authority', 'jurisdictional_variance', 'incomplete_context', 'ambiguous_holding', 'dicta_not_holding')",
      "description": "string (explanation of the uncertainty)"
    }
  ],
  "abstention": {
    "is_abstained": boolean,
    "reason": "string (if abstained, explain why the context is insufficient)"
  }
}

## CONSTRAINTS
[CONSTRAINTS]
- Every factual claim in the answer must be supported by at least one citation with a direct quote.
- Do not paraphrase holdings. Quote the exact language from the source.
- If sources conflict, surface the conflict in uncertainty_flags rather than choosing one.
- If the context does not contain enough information to answer, set abstention.is_abstained to true and explain what is missing.
- Do not cite sources that do not appear in the provided [CONTEXT].
- Distinguish between binding authority, persuasive authority, and dicta in your relevance explanations.
- If [RISK_LEVEL] is 'high', abstain unless the answer is unambiguously supported by binding authority in the user's jurisdiction.

## RISK_LEVEL
[RISK_LEVEL]

To adapt this template, start by mapping your retrieval pipeline's output into the [CONTEXT] placeholder. Each chunk should include a unique source_id, the full source name, jurisdiction metadata, and the raw text. The [CONSTRAINTS] placeholder lets you add domain-specific rules, such as 'prefer Supreme Court precedent over Circuit splits' or 'flag any source older than 10 years.' The [RISK_LEVEL] placeholder accepts values like 'low', 'medium', or 'high' and controls the abstention threshold. Before deploying, run this prompt through a citation hallucination eval: for each generated citation, verify that the direct_quote appears verbatim in the source chunk and that the source_id matches a real retrieved document. If your application operates in a regulated legal context, always route outputs through human review before they reach end users.

Integrating this prompt into a production legal RAG system requires more than a single API call. The harness must manage retrieval, prompt assembly, output validation, and a review workflow that reflects the high-stakes nature of legal content. Start by structuring the application as a pipeline: (1) user question intake with jurisdiction and date scoping, (2) retrieval from a vetted legal document store, (3) prompt assembly with the template below, (4) model inference, (5) structured output parsing, (6) automated validation, and (7) conditional human review before the answer reaches the end user. Each stage should log its inputs and outputs for auditability, and the system must never present an unvalidated answer as authoritative legal guidance.

Retrieval and context assembly. Before the prompt fires, retrieve passages from a curated corpus of case law, statutes, regulations, and contracts. Use a hybrid retrieval strategy—dense embeddings for semantic recall combined with keyword or boolean filters for jurisdiction, court level, and effective-date ranges. Deduplicate passages and rank them by a composite score that weights relevance, authority (e.g., binding precedent over persuasive), and recency. The top k passages become the [RETRIEVED_CONTEXT] placeholder. Include metadata per passage: source title, citation, court or jurisdiction, date, and a unique chunk identifier. This metadata is essential for the citation format the prompt demands and for downstream validation that checks whether every cited source actually appeared in the retrieved set.

Validation and retry logic. After the model returns an answer, parse it into a structured schema with fields for answer_text, citations (array of objects with source_id, quote, relevance), uncertainty_flags (array of strings), and abstention (boolean). Run automated evals before any human sees the output: (a) citation grounding—verify every source_id exists in the retrieved context and that quoted strings appear verbatim or with minimal, justifiable edits; (b) hallucination check—use a separate LLM judge prompt to flag claims not supported by the cited passages; (c) format compliance—confirm the output matches the expected schema and citation style. If validation fails, retry with the same prompt and an added [VALIDATION_ERRORS] block describing what went wrong. Cap retries at two attempts; after that, escalate to a human review queue with the full trace attached.

Human review and deployment considerations. For legal workflows, a human-in-the-loop gate is non-negotiable. Route answers that pass automated validation to a review interface that displays the question, the generated answer, the cited sources side-by-side with the model's quotes, and any uncertainty flags. The reviewer should confirm or override the answer, add disclaimers, and approve or reject before the answer is surfaced. Log the reviewer's identity, timestamp, and action for the audit trail. Model choice matters: use a model with strong instruction-following and long-context handling (e.g., GPT-4o, Claude 3.5 Sonnet). Set temperature low (0.0–0.2) to reduce variance in citation formatting. Monitor production for citation hallucination rate, abstention rate, and reviewer override frequency—these metrics will tell you when the prompt or retrieval pipeline needs tuning.