moorcheh/mair-floating-point-vs-quantized-benchmarks

Dataset Coverage

The benchmarks evaluate performance across 14+ specialized datasets covering:

• Legal & Regulatory: ACORDAR, AILA2019-Case, AILA2019-Statutes, LeCaRDv2, LegalQuAD, REGIR-EU2UK, REGIR-UK2EU.
• Financial: ConvFinQA, FinanceBench, FinQA, FiQA, HC3Finance.
• Medical & Clinical: NFCorpus.
• General/API: Apple Documentation.

Metrics

1. NDCG@10: Normalized Discounted Cumulative Gain at rank 10, measuring retrieval quality.
2. Latency (ms): Mean search latency measured on the server-side and end-to-end.

Benchmark Configurations

The data is organized into four primary splits, which you can switch between using the "Viewer" tab above:

1. NDCG@10 - Floating Point (ndcg_vector_fp): Evaluation of retrieval accuracy using standard 32-bit floating-point vectors.

2. NDCG@10 – Floating Point (Categorized) (ndcg_vector_fp_categorized): Evaluation of retrieval accuracy using standard 32-bit floating-point vectors, reported per dataset category.

3. Latency - Floating Point (latency_vector_fp): Search performance (ms) for floating-point vector retrieval.
4. NDCG@10 - Quantized (ndcg_vector_quantized): Evaluation of retrieval accuracy using optimized quantized/binary embedding paths.

5. NDCG@10 – Quantized (Categorized) (ndcg_vector_quantized_categorized): Evaluation of retrieval accuracy using optimized quantized or binary vector representations, reported per dataset category.

6. Latency - Quantized (latency_vector_quantized): Search performance (ms) for optimized quantized vector retrieval.

Tested Providers

Results include comparisons across:

• Moorcheh
• Elasticsearch
• Pinecone (with Cohere Rerank)
• PGVector (PostgreSQL)
• Qdrant

How to use in Python

You can load these results directly into a Pandas DataFrame using the Hugging Face datasets library:

from datasets import load_dataset

# Load the Latency results for Floating Point vectors
dataset = load_dataset("moorcheh/Benchmarks", split="latency_vector_fp")
df = dataset.to_pandas()

print(df.head())

Citation

If you use this dataset, please cite:

• arXiv:2601.11557