Elara’s predecessor had gone mad trying to manually set distance thresholds. He’d left a final note in the codebase: “You can’t teach a god where to stop listening.”
This vector was a concept with no neighbors. A data point with no context. weaviate autocut
The core mechanism of autocut is based on . When you perform a vector search or hybrid search , Weaviate calculates a similarity score for every candidate object. Often, these scores will have small differences among the top results but then show a significant "jump" or drop-off where the relevance decreases sharply. Elara’s predecessor had gone mad trying to manually
Imagine a nearText search in Weaviate that returns six objects with the following vector distances: [0.189, 0.190, 0.191, 0.250, 0.255, 0.320] . weaviate autocut