179 questions
No questions match those filters.
← All questions
RAGHard· GNN· Data Drift· System Design
Your GNN-RAG retriever’s accuracy quietly drops three m...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansIf the code has not changed, the input data—the graph—has likely shifted. A new relation type introduced to the graph is exactly the input the GNN was not trained to handle. You should check if the relation vocabulary has moved. The solution is to retrain the model on the new schema or implement a routing mechanism that sends entities reachable only through these new relations to a symbolic fallback, verifying if the fallback is carrying more of the answer set than before.