Tuning Recall and Latency

Vector search is approximate: it probes a subset of the index's vector clusters rather than scanning every vector. Probing more clusters improves recall (i.e. how often the true nearest neighbors are returned) at the cost of higher latency. Two session-level settings control this tradeoff:

<ParamField body="paradedb.vector_cluster_probe_epsilon" default={0.5}> The primary recall driver. A higher epsilon improves recall by probing more clusters at the expense of higher latency. Think of epsilon as an early termination value, where lower epsilon terminates early more aggressively. When tuning this value, we recommend sweeping in increments of `0.1`. Must be between `0.0` and `100.0`. </ParamField> <ParamField body="paradedb.vector_cluster_max_probe" default={0.02}> A control for tail latency. This setting enforces a ceiling on how many clusters a query may probe. For instance, at the default value of `0.02`, at most 2% of clusters are probed. Must be between `0.000001` and `1.0`, where `1.0` effectively means "no ceiling."

Note that setting this value to 1.0 does not equal exhaustive search, since the epsilon value still drives early termination.

</ParamField> <Note> Recall also depends on the index's `centroid_ratio`, set at build time. See [Index Options](/documentation/indexing/indexing-vectors#index-options). </Note>