Module 03 Summary

What this module established

Embeddings measure aboutness. That is why "money back" retrieves a refund policy without sharing a word, and equally why two chunks stating opposite policies score nearly the same - negation is not a property the geometry represents. Retrieval always returns k results, whether or not anything relevant exists.

Carry forward

• The rank-1-to-rank-2 score gap tells you whether the ranking was decisive or effectively a coin flip.
• k is chosen by measuring recall on your own eval set and stopping where the curve flattens, not by copying a default.
• Filters run before ranking. Post-filtering wastes top-k slots and means a restricted chunk was read before being dropped.

Before moving on

Move on when you can produce a per-question table of top score, score gap and correct-chunk rank for your own corpus, and defend your choice of k from it.

Mark complete