tramy-encoder

This is a fine-tuned alphaedge-ai/multilingual-e5-small-vie-32768 embedding model, specially optimized for Vietnamese Tourist & Q&A Assistant Retrieval with vector databases such as Qdrant.

Model Details

• Base model: alphaedge-ai/multilingual-e5-small-vie-32768 (Vietnamese-trimmed vocabulary 32,768 tokens, ~137MB)
• Language: Vietnamese (vi), English (en)
• Embedding dimension: 384
• Primary use case: Asymmetric retrieval (User search queries -> Stored Instructions/Documents/Passages)

⚠️ Important Prefix Usage (E5 Architecture)

Because this model uses the E5 architecture, you must prepend prefixes:

• When storing in Qdrant (Documents / Instructions / Answers): Prepend passage:
  • Example: passage: Thông tin tuyến xe buýt ở hồ chí minh
• When searching in Qdrant (User queries): Prepend query:
  • Example: query: tìm tuyến xe buýt ở hồ chí minh

Usage with sentence-transformers

from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity

# Load model (384 dimensions)
model = SentenceTransformer('lmtri0312/tramy-encoder')

# 1. Encoding documents to store in Qdrant
docs = [
    "Thông tin tuyến xe buýt ở hồ chí minh",
    "Lịch trình du lịch Đà Lạt 3 ngày 2 đêm"
]
doc_embeddings = model.encode([f"passage: {d}" for d in docs], normalize_embeddings=True)

# 2. Encoding user search query
query = "tìm tuyến xe buýt ở hồ chí minh"
query_embedding = model.encode([f"query: {query}"], normalize_embeddings=True)

# 3. Calculate cosine similarity
similarity = cosine_similarity(query_embedding, doc_embeddings)
print("Similarity:", similarity)