Skip to content

Exclude duplicate-vector self edges from the PaCMAP projection graph #615

Closed

Description

@realmarcin

Independent review of PR #613 found an exercised duplicate-vector edge case in the pinned PaCMAP 0.9.1 FAISS neighbor construction used by embedding_compute.projection.

The dependency excludes self by dropping result column zero (nbrs = I[:, 1:]). Identical query vectors share nearest-result ordering, so column zero cannot identify self for every member of a duplicate group. Its subsequent pair sampler accepts the resulting self edges, displacing real neighbor edges in the projection optimization graph.

The retained corpus contains 99 duplicate vector rows from 45 duplicate input-text groups (largest group: 11 rows). The public original-vector cosine neighbor list correctly excludes self; the affected component is the PaCMAP optimization graph.

Construct and validate neighbor pairs with identity-based self exclusion before projection. Preserve every discussion point, test duplicate vectors, and regenerate the projection from verified existing vectors without rerunning model inference.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions