Practice question from Foundations of Modern AI Search
A company indexes 1 million documents using text-embedding-ada-002. Six months later, they switch to all-MiniLM-L6-v2 for query encoding to save costs. Users report search quality has dramatically declined, but there are no errors. What explains this failure? A. The query embedding model has lower dimensionality than the document model B. Different embedding models create incompatible vector spaces where similarity comparisons are meaningless C. The new model processes queries too quickly, causing synchronization issues D. Open-source models cannot match commercial model quality for production use E. The vector database index needs recomputation after switching embedding providers
Answer
B
Explanation
Each embedding model learns its own way of organizing semantic meaning in vector space. Using one model for documents and another for queries is like using coordinates from different maps - the similarity comparisons become meaningless because concepts are positioned differently in each model's space. The system must use the same model for both indexing and querying. Option A is wrong because dimensionality mismatch would cause technical errors, not silent quality degradation. Option E is close but misses the fundamental issue that you cannot mix models at all - you must re-embed all documents, not just recompute the index.