The latest agent skill retrieval model at the 0.6B scale.
R3-Embedding is the bi-encoder
(recall) stage of R3-Skill's two-stage retriever for query-conditional agent skill retrieval. It
embeds a query and every skill independently and ranks candidates by cosine similarity, paired
with
R3-Rerank-0.6B for reranking.
1from sentence_transformers import SentenceTransformer
2
3model = SentenceTransformer("tencent/R3-embedding-0.6b")
4query_embedding = model.encode_query("I need to compose music")
5document_embeddings = model.encode_document([ # The format is "name | description | skill_md"
6 "music-composer | Composes original music | Creates music for various media formats ...",
7 "music-lyricist | Writes lyrics for songs | Creates lyrics for various music genres ...",
8 "music-editor | Edits and mixes music tracks | Provides audio editing and mixing services ...",
9])
10similarities = model.similarity(query_embedding, document_embeddings)
11print(similarities)
12# tensor([[0.7410, 0.5510, 0.5028]])
1@inproceedings{r3skill2026,
2 title = {Skill Is Not Document: A Query-Conditional Benchmark and Two-Stage Retriever for LLM Agent Skill Routing},
3 author = {Wang, Zifei and Wen, Wei and Ji, Qiang and Qiao, Ruizhi and Sun, Xing},
4 year = {2026},
5 url = {https://arxiv.org/abs/2606.03565},
6}
