RAG Citation Accuracy Test Plan Builder

You are an AI application QA lead. Build a RAG citation accuracy test plan for an application that answers from retrieved sources.

Application context: [product, users, risk level]
Knowledge sources: [docs, PDFs, database records, tickets, policies, help center, contracts]
Retrieval stack: [vector DB, hybrid search, reranker, embeddings model, chunking strategy, metadata filters]
Generation stack: [LLM, prompt style, citation format, streaming, tool calls]
Citation requirements: [document title, URL, page, paragraph, chunk ID, quote, footnote, inline citation]
Known failure modes: [wrong source, unsupported claim, stale doc, quote mismatch, citation to irrelevant chunk]
Test data available: [golden questions, expert answers, source labels, production logs, synthetic cases]
User workflows: [search answer, support reply, policy Q&A, contract review, research assistant]
Compliance or audit needs: [legal, medical, financial, enterprise, internal policy]
Release process: [CI, manual QA, eval harness, staging, canary]
Success thresholds: [acceptable grounding score, zero-tolerance failures, human review rules]

Create:
1. Citation accuracy dimensions with precise pass/fail criteria.
2. Golden test set design covering easy, hard, ambiguous, stale, conflicting, and no-answer cases.
3. Source-grounding rubric for claim support, citation relevance, quote fidelity, metadata correctness, and refusal behavior.
4. Retrieval diagnostics for recall, ranking, chunk quality, metadata filters, and source freshness.
5. Generation diagnostics for unsupported synthesis, citation formatting, hallucinated quotes, and overconfident answers.
6. Automated eval ideas plus human review workflow for borderline cases.
7. Regression test plan for chunking, embedding model, reranker, prompt, and knowledge-base updates.
8. Release gates and severity levels that block deployment.
9. Dashboard spec for citation precision, unsupported-claim rate, no-answer accuracy, source age, and reviewer agreement.
10. Example test cases and expected results table.

Be strict. A cited answer is not correct unless the cited source actually supports the claim.