Field MinCalibratedRelevance
The default relevance floor for surfaces that calibrate to the embedding model's own scale, expressed so that 0 is the model's similarity between unrelated texts and 1 is identity.
public const float MinCalibratedRelevance = 0.2
Returns
- float
- The default relevance floor for surfaces that calibrate to the embedding model's own scale, expressed so that 0 is the model's similarity between unrelated texts and 1 is identity.
Remarks
Prefer this over MinRelevanceScore wherever the model in use is known, as it holds its meaning across model families instead of behaving like a different filter on each. The value keeps the large majority of correct matches while rejecting almost every query the store cannot answer, measured across the supported embedder families.