179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsMedium· Multimodal Retrieval· CLIP· Thresholding
Your multimodal retriever never returns any images, eve...
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 plansA strong diagnostic approach begins by pulling the raw image similarity scores for a known-good query to compare them against the system’s cutoff. It is common for global thresholds, often calibrated on text, to be set significantly higher than the actual scoring range of a CLIP-based image encoder. A weak answer would prematurely suggest retraining the image encoder without first verifying the score distribution produced by the current model.