179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsVery Hard· System Design· Growth Strategy· Quantization
Your team wants 8-bit scalar quantization to cut the RA...
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 plansHNSW-Flat at 3,332 bytes per vector is 666 GB. Using 8-bit scalar quantization reduces this to 1,028 bytes (206 GB), and switching to IVF reduces it further to 776 bytes (155 GB). However, 10% monthly growth is 3.14x per year, meaning a 4.3x gain only buys 15 months before you are back in the same situation. A constant factor cannot solve exponential growth. The decision is not quantization versus more RAM, but whether to shard across nodes, using quantization as a bridge to buy time for that architectural transition.