Module 05 Summary

What this module established

A vector store always returns k results, relevant or not - so refusal requires scores and a threshold you set from your own distribution. Exact conditions belong in filters applied before ranking; similarity turns a hard guarantee into a soft preference.

Carry forward

• Everything past the answer-bearing chunk is noise paid for on every query. Measure recall and rank, and stop where the curve flattens.
• Compare retrievers on identical cases with one variable changed, storing results per case. A total hides a change that fixed one and broke another.
• Derive chunk ids from document and position. Positional ids break on rebuild; content hashes break when a typo is fixed.

Before moving on

Move on when your k is justified by measured recall, filters run before ranking, and your ids survive a rebuild that reorders the corpus.

Mark complete