<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> KURE is a family of Korean-English retrieval embedding models from Korea University's NLP & AI Lab and HIAI Institute. It addresses the need for high-performance bilingual information retrieval, offering dense, sparse, and reranker models. This suite benefits researchers and developers working with mixed-language or Korean-centric datasets.

How It Works

The KURE family includes dense (single-vector) models like KURE-v3 and KoE5, learned sparse retrieval (LSR) models such as KURE-Sparse, and cross-encoder rerankers like KURE-Reranker. Dense models generate vector embeddings for similarity search. Sparse models output vocabulary IDs and weights for efficient retrieval. Rerankers refine initial search results. KURE-v2 employs a distinct late-interaction, multi-vector architecture.

Quick Start & Requirements

Environment management is handled via uv (uv sync). API access for KURE-v3, KURE-Sparse, and KURE-Reranker-nano is available at kure.kullm.ai (requires GitHub API key). Local usage examples leverage sentence-transformers. sentence-transformers>=6.0 is required for KURE-v2. No explicit GPU/CUDA requirements are listed for basic setup.

Highlighted Details

• Bilingual Performance: Optimized for Korean and English retrieval tasks, achieving strong results on Korean benchmarks.
• Model Diversity: Offers dense, sparse, and reranker models for flexible retrieval pipeline construction.
• Benchmarking: KURE-v2 leads late-interaction models with Avg nDCG@10 of 0.8160 on Korean retrieval. KURE-v3 and KURE-Sparse-base also demonstrate competitive performance.
• Matryoshka Support: KURE-v3 supports variable vector dimensions (256-2048) for efficient storage and retrieval.
• API Service: Key models are accessible via a hosted API endpoint.

Maintenance & Community

Developed by Korea University's NLP & AI Lab and HIAI Institute. No specific community channels (e.g., Discord, Slack) or public roadmaps are detailed in the README.

Licensing & Compatibility

Released under the MIT License, permitting broad use, including commercial applications and integration into closed-source projects.

Limitations & Caveats

The API has request limits (8 texts for embeddings, 100 documents for reranking). The README does not detail specific limitations, known bugs, or unsupported platforms.