Module 09 Activity

Scenario

Embeddings make image search possible and make several new mistakes possible too.

What you build

A working image search over your own images, with a near-duplicate audit and a record of which model built the index.

Steps

1. Embed your images, normalise once, and build a search index. Record which model produced it.
2. Run five queries and inspect what the top matches have in common. Report whether it is the object or the background.
3. Report the rank-1-to-rank-2 gap for each query and say which matches were decisive.
4. Query for something absent from your corpus. Confirm results come back anyway and set a score threshold.
5. Run a near-duplicate check at cosine > 0.95 across all pairs, BEFORE splitting.
6. Report how many near-duplicate pairs would have straddled your split.

Evidence to hand in

• The index with the model recorded.
• Five queries with what the top matches share.
• Rank gaps per query.
• The absent-object query and your chosen threshold.
• The near-duplicate pairs and how many would have straddled the split.

Review checklist

• The embedding model is recorded with the index.
• The near-neighbour audit reports what actually drove the matches.
• A score threshold exists, derived from a measured distribution.
• The duplicate check ran before splitting.
• File hashes were not used as the duplicate check.

Mark complete