nlpai-lab/KURE-v2

<a href="https://github.com/nlpai-lab/KURE"> <img src="kure_logo.png" width="50%"/> </a>

🔎 KURE-v2

KURE-v2 is a Korean-English bilingual late-interaction (multi-vector) retrieval model built on skt/A.X-Encoder-base. It encodes every token into a 128-dimensional vector and scores query–document pairs with MaxSim.

At 154M parameters it averages nDCG@10 0.8160 across the nine MTEB(kor, v2) retrieval tasks — the strongest model on that suite, above every single-vector model measured, including one 175× its size.

It is trained in two stages: unsupervised contrastive pretraining on 20.7M unlabeled pairs (KURE-v2-unsupervised), then supervised fine-tuning with contrastive loss and KL distillation from a reranker.

Key Characteristics

• —Late interaction: one 128-d vector per token, scored with MaxSim.
• —Long documents: up to 8,192 tokens.
• —Compact: 154M parameters.
• —No instruction prefixes: queries and documents need no task instruction. Query expansion to 64 tokens is handled by the model.

Usage

<details> <summary><b>PyLate Usage</b></summary>

pip install -U pylate

Indexing documents

from pylate import indexes, models, retrieve

model = models.ColBERT(model_name_or_path="nlpai-lab/KURE-v2")

index = indexes.PLAID(
    index_folder="pylate-index",
    index_name="index",
    override=True,
)

documents_ids = ["1", "2", "3"]
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,
    show_progress_bar=True,
)

index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

To reuse an existing index, instantiate it without override:

index = indexes.PLAID(index_folder="pylate-index", index_name="index")

Retrieving top-k documents

retriever = retrieve.ColBERT(index=index)

queries_embeddings = model.encode(
    ["훈민정음은 언제 만들어졌나요?"],
    batch_size=32,
    is_query=True,
    show_progress_bar=True,
)

scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,
)

Reranking

To rerank a first-stage candidate list without building an index:

from pylate import models, rank

model = models.ColBERT(model_name_or_path="nlpai-lab/KURE-v2")

queries = [
    "전기차 폐배터리는 어떻게 재활용하나요?",
    "겨울에 한라산을 오를 때 필요한 장비는?",
]
documents = [
    [
        "폐배터리에서 리튬과 코발트를 회수하는 습식 제련 공정이 상용화되고 있다.",
        "급속 충전기는 30분 내외로 배터리를 80%까지 충전할 수 있다.",
    ],
    [
        "겨울 한라산 산행에는 아이젠과 방한 장갑이 필수이며 입산 시간이 제한된다.",
        "제주 올레길은 해안을 따라 이어지는 27개 코스로 구성되어 있다.",
        "적설기에는 등산화에 스패츠를 착용해 눈이 들어가는 것을 막는 것이 좋다.",
    ],
]
documents_ids = [[1, 2], [1, 3, 2]]

queries_embeddings = model.encode(queries, is_query=True)
documents_embeddings = model.encode(documents, is_query=False)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

</details>

<details> <summary><b>Sentence-Transformers Usage</b></summary>

pip install -U sentence-transformers

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("nlpai-lab/KURE-v2")

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings[0].shape)
# (64, 128) (29, 128)

# MaxSim late-interaction scoring (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)

</details>

Evaluation

KURE-v2 is evaluated on the nine MTEB(kor, v2) Retrieval tasks. We report nDCG@10. The results are also shown in the official MTEB Leaderboard?types=Retrieval&s.summary=meanTask&d.summary=desc)

Late-interaction rows were measured with mteb 2.18.16 and PLAID retrieval. Single-vector rows are taken from the official MTEB results repository, except for Belebele, where only the Korean-query / Korean-corpus subset is used. The original version also includes cross-lingual subsets (Korean query – English corpus, English query – Korean corpus).

Serving

KURE-v2 is a late-interaction model: each document is stored as a set of token vectors, so the practical questions for deployment are index size and search cost. We benchmarked KURE-v2 across ANN backends and compression schemes on the 9 Korean MTEB retrieval tasks, against five single-vector baselines served with faiss HNSW. All numbers are end-to-end: batch-1 query encoding + index search, measured serially on one A100 80GB.

<p align="center"> <img src="assets/deploy_overview.png" width="100%" alt="Average nDCG@10 vs. index storage (left) and vs. end-to-end QPS (right)"> </p>

Two things the figures show:

• —Hierarchical token pooling (x2) halves the index for a 0.04 nDCG drop. Asymmetric binary quantization (1-bit document tokens, bf16 queries) shrinks it 9.4x for 1.05. Stacking the two (pooling x3 + binary), the entire 9-corpus index fits in 1.7 GB, smaller than every single-vector HNSW index (13.1-50.0 GB), while still outscoring the best single-vector model (79.57 vs 79.07).
• —A live query arrives as text: 4B-8B single-vector models spend 38-40 ms encoding it, capping them at ~25 QPS no matter how fast HNSW is. KURE-v2 encodes in 13.8 ms (154M params), so every configuration except MUVERA serves 43-55 QPS, roughly 2x the 8B single-vector models, at higher quality.

Large corpora: tail latency

<p align="center"> <img src="assets/bigcorpus_miracl.png" width="70%" alt="MIRACL (1.5M docs): quality, e2e p95 latency, index size"> </p>

On the largest corpus (MIRACL, ~1.5M documents) an exhaustive 1-bit scan costs O(corpus): p95 climbs to 156 ms, and pooling the tokens 3x only brings it to 74 ms. Generating candidates with faiss BinaryIVF (Hamming search over the same 1-bit index) and re-scoring them with exact asymmetric MaxSim cuts p95 to 38 ms on the same 2.2 GB index, lower tail latency than the 4B-8B single-vector baselines (43 ms) at higher nDCG. For large collections, use a candidate-generating index (PLAID or BinaryIVF), not an exhaustive scan.

<details> <summary><b>Measurement details</b></summary>

• —Hardware: 1x NVIDIA A100 80GB, 2x AMD EPYC 7513 (64 cores), 1.2 TB RAM.
• —Software: faiss-cpu 1.15.0, fast-plaid 1.6.0, sentence-transformers 6.0.0, PyTorch 2.8.0.
• —Protocol: batch-1, serial. Index-search latency: 10 warmup queries, then every query of the task measured once (QPS = 1/mean). Query-encoding latency: 5 warmup, 50 measured. End-to-end = encoding + search.
• —Precision: encoding in bf16; each index stores its own format (HNSW fp32, PLAID 4-bit residuals, binary 1-bit).
• —Index size: the full serialized index on disk (vectors, graph, codebooks; external doc-id mapping excluded).
• —Tasks: the 9 Korean MTEB retrieval tasks; MLDR is the mean of its dev/test splits; nDCG@10 x100.
• —HNSW: IndexHNSWFlat (inner product on L2-normalized embeddings), M=32, efConstruction=200, efSearch=64.
• —PLAID: nbits=4, all other settings fast-plaid defaults (kmeansniters=4, nivfprobe=8, nfull_scores=4096). nbits=2/1 give 27.0/17.0 GB at 81.25/81.09 nDCG.
• —MUVERA: numrepetitions=10, numsimhashprojections=6, finalprojection_dimension=8192, exact-MaxSim rerank of the top 1,000.
• —BinaryIVF: nlist=floor(sqrt(total tokens)) capped at 65,536, nprobe=32, top-128 Hamming tokens per query token, exact asymmetric-MaxSim rerank of the top 1,000 documents.
• —Token pooling: hierarchical (Ward linkage), pool_factor 2-3, documents only. </details>

Citation

@misc{kure-v2,
  title  = {KURE-v2: a Korean-English bilingual late-interaction retriever},
  author = {Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-v2},
}

@inproceedings{jang2025kure,
  title={KURE: Embedding Model for Korean-Specific Retrieval},
  author={Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Park, JeongBae and Lim, Heuiseok},
  booktitle={Annual Conference on Human and Language Technology},
  pages={129--134},
  year={2025},
  organization={Human and Language Technology}
}

@inproceedings{santhanam-etal-2022-colbertv2,
  title     = {ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction},
  author    = {Santhanam, Keshav and Khattab, Omar and Saad-Falcon, Jon and Potts, Christopher and Zaharia, Matei},
  booktitle = {Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year      = {2022},
  pages     = {3715--3734},
}

@misc{PyLate,
  title  = {PyLate: Flexible Training and Retrieval for Late Interaction Models},
  author = {Chaffin, Antoine and Sourty, Raphaël},
  year   = {2024},
  url    = {https://github.com/lightonai/pylate},
}