agentrag-embed-v1

Vietnamese-medical fine-tune of intfloat/multilingual-e5-base for RAG retrieval in the AgentRag project.

• Pooling: mean (+ L2 normalize)
• Dimensions: 768
• Max sequence length: 512
• Domain: Vietnamese medical documents
• Training: 5.3k (query, positive, negative) triplets, MultipleNegativesRankingLoss, 2 epochs
• Eval: recall@10 +0.20 over base e5 on the project's retrieval benchmark (C1)

Serving with TEI

services:
  tei:
    image: ghcr.io/huggingface/text-embeddings-inference:cuda-latest
    command:
      - --model-id=dung6903/agentrag-embed-v1
      - --pooling=mean

Query prefix conventions follow e5: query: ... / passage: ....