RETRIEVAL DATA PRODUCT SPEC (ONE-PAGE TEMPLATE)

1) Name + purpose
- Product name:
- Intended users (roles/teams/tenants):
- Primary jobs-to-be-done (3–5 bullets):
- Explicit non-goals (what it must refuse or redirect):

2) Sources of truth (registry)
For each source, record:
- System: (Confluence / Notion / GitHub / Google Drive / SharePoint / Zendesk / Salesforce / etc.)
- What topics it is authoritative for:
- Owner (person/team):
- Content type: (policy / runbook / tutorial / contract / ticket / PRD)
- Deletion/retention expectations:

3) Freshness + lifecycle
- Freshness requirement by topic: (e.g., “policy must reflect latest published version”)
- Sync method: (webhook/poll/manual)
- Update cadence:
- What triggers an emergency re-index:
- Stale-content behavior: (refuse / warn / fallback to human)

4) Permissions + access model
- Identity provider(s): (Okta / Entra ID / Google Workspace / etc.)
- Permission mechanism per source: (ACLs, groups, folder sharing)
- Retrieval-time enforcement approach:
 - Required user attributes (dept, region, tenant, employment status):
 - Required document attributes/tags:
- Audit logging requirements: what gets logged and where it’s stored

5) Indexing & representation decisions
- Chunking policy (by heading/procedure/FAQ pairs/etc.):
- Required metadata fields (doc_id, source, owner, created_at, updated_at, sensitivity, region, product_area):
- Citation format required in outputs:
- Redaction rules (what must never appear in retrieved text):

6) Grounding rules (output constraints)
- “No citation, no claim” rule? (yes/no)
- Refusal categories (legal, HR, security, medical, financial, etc.):
- Escalation path (who/what to route to):
- Allowed tools/actions (if any) and hard stops:

7) Evals & release gates
- Golden question set location (repo/path):
- What gets tested on every change:
 - Retrieval correctness
 - Permission negatives
 - Citation presence/quality
 - Freshness after updates
 - Refusal behavior
- CI gate: what constitutes a fail:
- Human review requirements (who signs off):

8) Operational ownership
- DRI (directly responsible individual):
- On-call / escalation:
- Incident types and playbook link:
- Backlog for known issues:

9) Rollout plan
- First integration surface (Slack, web app, internal portal, IDE):
- Limited launch group:
- Feedback channel:
- Criteria to expand access:

Use this as a forcing function: if a field is hard to fill in, that’s the work. The model won’t save you from missing ownership, permissions, or freshness.