The best answer often depends on who is asking. A policy question from a manager in Perth may need a different answer from one asked by a contractor in London.
Signals for Personalisation
- Role, department and seniority.
- Location and applicable jurisdiction.
- Products or services the user has.
- Language.
- Past questions and preferences, where appropriate.
Where to Apply It
- Retrieval filters: only documents relevant to the user's region or product.
- Boosting: prefer content for their role.
- Prompt context: tell the model relevant user attributes so it frames the answer correctly.
- Tone and detail level.
Keep It Transparent
State the assumptions in answers: "For full-time staff in Australia, the policy is…". This lets users correct wrong assumptions.
Privacy and Fairness
Use only data the user would expect, avoid sensitive attributes, and ensure personalisation doesn't give some groups worse information.
Security
Personalisation data must not leak between users through caches, memory or logs.
Test Across Personas
Evaluation sets should include the same question asked by different personas, with the expected differences in answers.