RAG-07 · RAG prompt templates and safe prompt changes¶
Teams write RAG prompts from scratch on each project, so answer style and refusal behaviour differ across the portfolio. Prompt and model changes then go live without a regression test, and users find the quality drop first.
- Who reads it: GenAI engineers and prompt leads building and running RAG assistants, and DevOps engineers who own the release pipeline
- Outcome: Every prompt or model change passes an evaluation gate
- Pack owner: Deeparani (catalog owner role: Prompt Lead)
Master catalog entry¶
- Catalog name: RAG Prompt Template Pack (P2)
- Why it matters: Consistent answer style and refusal behaviour across the portfolio
- What it removes: Removes prompt-from-scratch work at the start of every RAG build
- Stack: Prompty, Jinja2, prompt registry
- Maturity target: Level 3 — grounded QA, summarisation, comparison, refusal, multi-doc
Playbooks¶
| ID | Story | Title | Kind | Complexity | Urgency | Status | Owner | Planned publish |
|---|---|---|---|---|---|---|---|---|
| PB-RAG-07-01 | RAG-07-S1 | Release a prompt change without regressions | Runbook | Medium | 5 | draft | Yasar | 2026-11-17 |
| PB-RAG-07-02 | RAG-07-S2 | Re-validate prompts on a model version upgrade or retirement | Runbook | Medium | 5 | draft | Gayathri Ramachandran | 2026-11-18 |
| PB-RAG-07-03 | RAG-07-S3 | Build the standard RAG templates: grounded QA, summary, comparison, refusal and multi-document | How-to | Low | 3 | draft | Gayathri Ramachandran | 2026-12-18 |
| PB-RAG-07-04 | RAG-07-S4 | Prompt injection and data-leak review checklist | Checklist | Low | 4 | draft | Gayathri Ramachandran | 2026-12-01 |