pub const EMBEDDING_LOAD_EXPECTED_RSS_MB: u64 = 1_100;Expand description
Expected RSS in MiB for a single instance with the ONNX model loaded via fastembed.
Used in the formula min(cpus, available_memory_mb / EMBEDDING_LOAD_EXPECTED_RSS_MB) * 0.5
to compute the dynamic permit count.
Value calibrated on 2026-04-23 with /usr/bin/time -v against sqlite-graphrag v1.0.3
on the heavy commands remember, recall, and hybrid-search, all peaking near
1.03 GiB RSS per process. The constant below rounds up with a defensive margin.