superlinked/external-benchmarking

Vector Search Benchmarks

This repo contains datasets for benchmarking vector search performance, to help Superlinked prioritize integration partners. For performing actual benchmarking on this dataset, see the github repository README.

Overview

We reviewed a number of publicly available datasets and noted 3 core problems + here is how this dataset fixes them:

Problems of other vector search benchmarks	How this dataset solves it
Not enough metadata of various types makes it hard to test filter performance	3 number, 1 categorical, 3 text, 1 image column
Vectors too small, while SOTA models usually output 2k+ even 4k+ dims	4154 dims
Dataset too small, especially if larger vectors are used	100k, 1M and 10M item variants, all sampled from the large dataset

Available Datasets

Product data

The data_dirs contain parquet files with the metadata and vectors.

Dataset	Records	# Files	Size
benchmark_10k	10,000	100	~230 MB
benchmark_100k	100,000	100	~2.3 GB
benchmark_1M	1,000,000	100	~23 GB
benchmark_10M	10,534,536	1000	~240 GB

The structure of the files is the same throughout:

Schema([('parent_asin', String), # the id
        ('main_category', String),
        ('title', String),
        ('average_rating', Float64),
        ('rating_number', Float64),
        ('description', String),
        ('price', Float64),
        ('categories', String),
        ('image_url', String)])
        ('value', List(Float64)), # the vectors

Data Access

The product metadata and vectors are available using HF Datasets.

from datasets import load_dataset

benchmark_10k = load_dataset("superlinked/external-benchmarking", data_dir="benchmark-10k")
benchmark_100k = load_dataset("superlinked/external-benchmarking", data_dir="benchmark-100k")
benchmark_1M = load_dataset("superlinked/external-benchmarking", data_dir="benchmark-1M")
benchmark_10M = load_dataset("superlinked/external-benchmarking", data_dir="benchmark-10M")

Dataset Production

Source Data

• Origin: Amazon Reviews 2023 dataset
• Categories: ["Books", "Automotive", "Tools and Home Improvement", "All Beauty", "Electronics", "Software", "Health and Household"]

Embeddings

The embeddings are created via a superlinked config. The resulting 4154 dim vector contains:

• 1 categorical,
• 3 number,
• 3 text (Qwen/Qwen3-Embedding-0.6B),
• and 1 image (laion/CLIP-ViT-H-14-laion2B-s32B-b79K)

embeddings concatenated.