Guide-linked quiz · Medium Level
Re-Indexing Embeddings When Models Change Quiz
Plan embedding migrations with versioned vector spaces, parallel indexes, backfills, relevance checks, safe cutover and rollback.
Question 1 of 8
Why should old and new model embeddings generally be kept in separate index versions?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- Why should old and new model embeddings generally be kept in separate index versions?
- Use an index migration instead of checking vector compatibility
- Assume equal vector dimensions mean matching semantic coordinates
- The two models may define different vector spaces even when dimensions match
- Store old query vectors with new candidate vectors because both are embeddings
- Which migration strategy keeps the current service available during a new embedding backfill?
- Delete the old index before generating replacement vectors
- Backfill a second index while the current index continues serving
- Switch the query encoder before rebuilding candidate vectors
- Mix both models’ vectors without storing their model versions
- Which check directly prevents a dimension error when writing new vectors?
- Confirm the input token count equals the index vector dimension
- Confirm query and document encoders have the same display name
- Confirm the training checkpoint dimension without checking index configuration
- Confirm each output vector dimension matches the configured index
- What does shadow querying help evaluate?
- Send old and new queries to separate evaluators without changing user-visible responses
- Compare only the raw text before either embedding model runs
- Route all live requests to the new index while retaining the old results
- Compare training gradients for the two embedding models
- Why reconcile updates that arrive during a long backfill?
- Change the query vector dimension during the backfill
- Skip query evaluation after vectors are written
- Remove the stable document identifiers before reconciliation
- Ensure source changes made during the scan are included before cutover
- What should a rollback restore after a failed migration?
- Restore the previous dashboard configuration only
- Restore a compatible old query encoder and old index pair
- Keep the new query encoder pointed at old vectors
- Roll back the unrelated training checkpoint
- Which evidence should guide an index cutover?
- The new model has lower training loss than the old model
- Vector generation completed for all records
- Index creation returned a ready status
- Check retrieval quality and operational metrics on representative queries
- What makes a backfill safer to retry?
- Use stable record IDs and idempotent writes
- Generate random IDs on every retry
- Ignore failed records and mark the scan complete
- Switch indexes after the first batch succeeds
Keep testing yourself
More quizzes picked for your level.