Skip to content

RAG-06 · Citations, grounding and no-answer handling

Citations and grounding decide whether users can trust a RAG assistant. In regulated settings, an assistant that cites loosely or answers beyond its sources cannot be defended. Teams also change prompts or models without measuring grounding, and often learn of a regression only when users report it.

The playbooks in this pack cover the full cycle: returning claim-level citations, deciding when the assistant should not answer, gating releases on groundedness, reviewing prompt and model changes, and investigating a wrong answer or wrong citation in production.

  • Who reads it: GenAI engineers building answer generation and DevOps engineers operating releases
  • Outcome: Every claim cited and grounding measured per release
  • Pack owner: Deeparani (catalog owner role: RAG Lead)

Master catalog entry

  • Catalog name: Citation & Grounding Framework (P1)
  • Why it matters: Citations are what make an assistant defensible in regulated settings
  • What it removes: Span attribution and citation rendering solved end to end
  • Stack: Python, Azure OpenAI structured outputs, React renderer
  • Maturity target: Level 3 — includes groundedness scoring and no-answer handling

Playbooks

ID Story Title Kind Complexity Urgency Status Owner Planned publish
PB-RAG-06-01 RAG-06-S1 Return claim-level citations with Azure OpenAI structured outputs How-to Medium 3 draft Deeparani 2026-12-03
PB-RAG-06-02 RAG-06-S2 Define no-answer and low-confidence handling Decision guide Medium 3 draft Deeparani 2026-11-30
PB-RAG-06-03 RAG-06-S3 Score groundedness and gate releases How-to Medium 4 draft Gayathri Ramachandran 2026-12-04
PB-RAG-06-04 RAG-06-S4 Investigate an ungrounded answer or wrong citation incident Runbook Medium 5 draft Deeparani 2026-11-12
PB-RAG-06-05 RAG-06-S5 Prompt or model change without grounding regressions checklist Checklist Low 4 draft Gayathri Ramachandran 2026-11-19