Cluster Parameters

The 'Min Relevance' property defines a hard threshold for semantic similarity. Chunks with a score below this value are always excluded from the cluster, regardless of their proximity or relative score. This helps filter out weak or irrelevant matches, ensuring that only content with sufficient semantic alignment is considered for extraction.

Increasing this value enforces stricter matching, while lowering it may allow more loosely related content to be included.

The 'Top N' property controls how many of the highest-scoring chunks are considered for inclusion in the cluster. Only the top N results, sorted by similarity score, are evaluated using the other clustering criteria.

Reducing this value can improve performance and restrict the cluster to the most relevant content. Increasing it allows a broader context to be considered, which may be useful for longer or more complex passages.

The 'Min Relative Score' property defines a relative threshold for inclusion in the cluster. A chunk must have a score at least this percentage of the top-scoring chunk to be eligible. This ensures that all included chunks are of comparable relevance, preventing the inclusion of marginally related content.

This property is only applicable when 'Top N' is greater than 1.

The 'Accept Threshold' property provides an override for highly relevant content. If a chunk's score exceeds this absolute threshold, it will be included in the cluster even if it does not meet the relative score requirement.

This is useful for ensuring that very strong matches are never excluded due to relative scoring rules. This property is only applicable when 'Top N' is greater than 1.

The 'Require Overlap' property enforces continuity within the cluster. When enabled, only chunks that overlap with at least one other accepted chunk (besides the top-scoring chunk) are included. This helps ensure that the resulting quote region is contiguous and avoids grouping together disjointed or scattered content.

Disabling this property may allow non-contiguous chunks to be clustered, which can be useful in rare cases where relevant content is separated by unrelated text. This property is only applicable when 'Top N' is greater than 1.