Module 10 Summary

The idea this module was built around

Embed once, cache with the model version, reuse across tasks. The expensive step should happen exactly once.

What you can now do

• Reuse cached vectors across several downstream tasks
• Choose between cosine and Euclidean, and know why normalising reconciles them
• Find intrinsic dimensionality with a bottleneck sweep
• Probe an embedding for a sensitive attribute

The trap this module removed

Assuming that excluding an attribute as a feature means the model cannot use it. Correlated features reconstruct it, and the probe is the only way to find out.

Figures worth remembering

Reconstruction error collapses at the true dimensionality. Vectors from two model versions are not comparable, even for the same input.

Before you move on

Run a leakage probe on any embedding you use for ranking or retrieval, once per sensitive attribute.

Mark complete