An AI result cache serves obsolete policy answers after the underlying documents change. Which design should be improved?

1. Invalidation or versioned cache keys tied to relevant source and configuration changes
2. Remove all source version information
3. Cache all answers indefinitely to maximize hits
4. Increase semantic similarity tolerance to reuse more old answers