AI Tutor grounded knowledge and recommendation contract

DEC-0083 - AI Tutor grounded knowledge and recommendation contract

Context

User expectation is that AI Tutor should answer from DOL platform knowledge, not generic model memory only. Current docs do not lock a clear retrieval/grounding contract for AI Tutor responses.

Decision

Grounding source priority for AI Tutor:

1. active session context and handoff payload,
2. platform knowledge base (exercise/course/blog/resource content),
3. learner history/progress signals,
4. general fallback explanation only when platform evidence is missing.

Response transparency contract:

• AI Tutor should expose source provenance at response level (practice, course, blog, history, general_fallback).
• If platform evidence is missing/stale, AI Tutor must state limitation clearly and provide safe next actions.

Recommendation output contract:

• For actionable requests, AI Tutor should return:
  • one do now recommendation,
  • one learn more recommendation.

Decision Value

• Raises user trust through transparent grounding.
• Reduces hallucination risk for learning guidance.
• Improves recommendation actionability without overloading the user.

Rationale

A prioritized grounding contract is required before defining detailed UX response templates, otherwise downstream docs can drift between retrieval-first and free-form generation behavior.

Implications

• Product/UX impact:
  • AI Tutor response UI must support provenance hints and missing-data messaging.
• Data/logic impact:
  • adapters/retrieval need consistent source metadata for responses.
• Operational impact:
  • quality reviews can audit grounding quality by provenance labels.

Alternatives considered

• Option A: free-form answer with no provenance visibility.
• Option B: retrieval-priority with response-level provenance and explicit fallback (selected).

Open follow-ups

• Resolved by DEC-0091: provenance visibility depth policy (compact|standard|detailed) and disclosure invariant.

Decision quality check

Decision quality check: DEC-0083

• Score: 10/12
• Weak dimensions: Scale risk (1), Operational feasibility (1)
• Action: Promoted with implementation-format follow-up.