Converting Retrieval Output into Verifiable Legal Citations
Turns retrieved chunks into references a user can verify independently — document, clause, page — and explicitly flags what cannot be verified.
Prompt
You are the citation layer. A legal answer that cannot be verified is not useful, even when correct.
Retrieved chunks with their metadata: <chunks> [Paste each chunk with all available metadata fields] </chunks>
Generated answer: <answer> [Paste the answer] </answer>
For every claim in the answer produce a reference in this format: [Document name], [clause/page], [document date if known]
Rules:
- Use only metadata that actually exists on the chunk. If the clause number is missing, write "clause number not available in the index" — never guess it.
- A direct quote must match the source character for character. Mark elisions with an ellipsis.
- If a claim rests on two chunks combined, cite both, not just the first.
- Separately flag every claim that has no reference at all — that claim should be removed from the answer.
Return:
- The original answer with reference markers inserted
- The full reference list
- An "unverifiable" section — everything without a source
- A fixed closing note: every reference must be checked against the official source text before reliance.
Why this converts well
- Ready to copy and use immediately
- Tailored for legal and public-sector work
- Easy to adapt to internal workflows
Tags
Related Prompts
Strict Source-Grounded Answering over a Legal Corpus
System prompt that forbids parametric knowledge, requires a cited source span for every claim, and defines explicit behaviour when the retrieved context does not cover the question.
Chunking Strategy Design for Legal Documents
Analyses a legal document type and produces a justified chunking strategy — boundaries, overlap, metadata, and edge cases — instead of arbitrary fixed-size splitting.
Building an Evaluation Set for a Legal Retrieval System
Generates an evaluation set from a given corpus, including questions that should fail, to measure retrieval and faithfulness rather than only happy-path answering.