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

yaml

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: ....