Definition
Plain language
The pre-built catalogue a search system flips through instead of re-reading everything.
As stated in the literature
The data structure over embeddings or terms that supports fast top-k lookup; what gets indexed for an image (its caption or its embedding) determines which manipulations are visible to retrieval and which are not.
Also called: search indexes
Why it matters: What you choose to put in the index determines what can ever be found — index only a picture's caption and any change the caption doesn't mention becomes unsearchable.
For example, a library catalogue lets you find a book by subject in seconds instead of walking every shelf.
Heard on the show
“Which means when the question comes in, the search index never files your image under the right name.”Episode 247 — One Edited Photo, an Honest Caption, and a RAG System That Believes It