VecLink

Cross-embedding vector correspondence via iterative geometric embedding hashing. Given two sets of embeddings (from different models) over the same data that are partially overlapping, VecLink identifies which vectors correspond to the same underlying entities using only a small set of seed anchors.

Installation

Requires Python 3.9–3.10.

# Install dependencies
uv sync

# PyTorch Geometric extensions (must be installed separately)
uv pip install torch-cluster torch-scatter torch-sparse torch-spline-conv \
  -f https://data.pyg.org/whl/torch-2.1.0+cu118.html

Quick Start

uv run veclink.py \
  --dataset scifact \
  --emb1 mistral \
  --emb2 openai \
  --overlap_ratio 0.3 \
  --n_seeds 15 \
  --seed 42 \
  --use_bernoulli_trials

Embeddings

Place embedding files in the embeddings/ directory as NumPy .npy files, named as:

corpus_embeddings_{model}_{dataset}.npy

For example: corpus_embeddings_mistral_scifact.npy, corpus_embeddings_openai_scifact.npy.

Key Arguments

Argument	Default	Description
--dataset	scifact	Dataset name
--emb1 / --emb2	mistral / openai	Embedding model names
--overlap_ratio	0.3	Fraction of data shared between the two sets
--n_seeds	None	Number of seed anchor pairs
--use_bernoulli_trials	True	Beta–Bernoulli posterior ensemble selection (paper default; pass False for the raw-vote baseline)
--max_iter	100	Maximum refinement iterations
--seed	None	Random seed for reproducibility
--use_gpu	True	Enable GPU acceleration

Supported Datasets

BEIR benchmarks: scifact, scidocs, fiqa, nfcorpus, arguana.