When Retrieval Lies — Distance, Similarity, and How to Debug It

1. What is the correct interpretation of a self-match score of `0.99` in a semantic retrieval system?

The model is faulty Higher scores indicate more similarity Lower scores indicate more similarity The vectors are corrupted

2. When encountering low-quality results in semantic search, what should the initial response be according to the lesson?

Change the embedding model Inspect the pipeline around the embedding model Reindex the vector store Completely redesign the collection architecture

3. What is the primary reason the initial diagnoses of the code search issue were incorrect?

The embedding model was too weak The results were incorrectly labeled Tests were run through the inverting code The vector storage format was outdated

4. What is the key strategy when encountering a suspected bug in a system according to the lesson?

Always reindex the data Isolate the variable with a controlled measurement Testing should be done in the same environment Focus only on end-to-end tests

Continue: The RAG Pattern → Review this lesson Retake quiz