- Find some related works on optimizing vector bases against noisy external processes (Adam)
- Do background research on how vector databases work (how they actually fetch entries and what part of this might be heavily dependent on memory availability and cache availability) (Adam)
- Measure cache contention versus latency (Marco)
- Note if this works
- Adjust memory measurement versus latency (Vivek)
- Note if this works
- Explore some optimizations
- ef_search adaptive (Vivek)
- quantization of vectors (Vivek)
- two types of collections: (1) compressed storage versus (2) uncompressed storage
- Switch to fetching from compressed (low dimensionality / quantized) vector db when there is memory pressure (maybe future work?)
- Test everything once polished on bare-metal Linux server
- Improvements on latency when there is cache and/or memory pressure