Calibrated admission thresholds for the built-in embedders, kept identical to the Python library.
Measured on a labelled set of 48 relevant and 528 unrelated query and memory pairs (2026-09-22).
The lexical embedder only matches shared words, so no threshold makes it semantic: 0.30 is where
it still recalls about 42% of relevant memories while admitting about 7% of unrelated ones.
bge-small-en-v1.5 at 0.63 recalls about 88% with about 84% precision, admitting 1.5% of unrelated
pairs. The old single default (0.7) recalled 4% with the lexical embedder and 71% with bge-small.
Calibrated admission thresholds for the built-in embedders, kept identical to the Python library. Measured on a labelled set of 48 relevant and 528 unrelated query and memory pairs (2026-09-22). The lexical embedder only matches shared words, so no threshold makes it semantic: 0.30 is where it still recalls about 42% of relevant memories while admitting about 7% of unrelated ones. bge-small-en-v1.5 at 0.63 recalls about 88% with about 84% precision, admitting 1.5% of unrelated pairs. The old single default (0.7) recalled 4% with the lexical embedder and 71% with bge-small.