Captured at (local ISO): 2026-05-18 05:17:26

1. Overview

GTE (General Text Embedding) is a text-embedding family from Alibaba DAMO / Tongyi Lab. On C-MTEB (Chinese MTEB), GTE models routinely rank near the top.

GTE-Base-ZH traits:

• Balanced: Base size hits a sweet spot between cost and recall.
• Broad: Search, recs, QA alignment, long-text clustering—typical RAG paths.
• Deployable: Low VRAM—consumer GPUs or even CPU.

2. Environment setup

Python ≥ 3.8. On a cloud image with PyTorch 2.9 already baked in, you can skip the torch install.

pip install torch transformers 

Then sentence-transformers:

pip install -U sentence-transformers

Third: tf-keras

pip install tf-keras

3. Smoke test script

import os
import warnings
# 必须在导入 sentence_transformers 之前执行
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", message=".*TRANSFORMERS_CACHE.*")

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'


os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'


os.environ['PYTHONWARNINGS'] = 'ignore'


from sentence_transformers import SentenceTransformer


warnings.filterwarnings("ignore", category=FutureWarning)

# 你的业务逻辑
model_name = 'thenlper/gte-base-zh' 
model = SentenceTransformer(model_name)

sentences = ["如何部署大语言模型", "深度学习入门指南"]
embeddings = model.encode(sentences)

print(embeddings.shape)

Save as .py and run:

If you see No module named 'transformers.modeling layers' (TensorFlow stack glitch):

pip install --upgrade transformers peft sentence-transformers

Healthy run looks like:

4. Example: semantic search matching

import os
import warnings

warnings.filterwarnings("ignore", category=FutureWarning)

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
os.environ['USE_TORCH'] = '1'

from sentence_transformers import SentenceTransformer, util

# 1. 加载模型 (自动识别你安装好的路径或从国内镜像下载)
print("正在加载 GTE 模型...")
model = SentenceTransformer('thenlper/gte-base-zh')

# 2. 准备数据：一个查询问题，和几个候选答案
query = "怎么才能让身体更健康？"

documents = [
    "保持充足的睡眠，每天坚持锻炼身体，多吃蔬菜水果。",  # 相关
    "健康是革命的本钱，合理的饮食结构对人体至关重要。",  # 相关
    "今天的天气真不错，非常适合出门郊游和野餐。",      # 弱相关 (有出门运动暗示)
    "Python是一种广泛使用的编程语言，适合开发AI模型。",   # 完全无关
    "由于全球气候变暖，北极熊的生存环境受到了威胁。"     # 完全无关
]

print(f"\n用户问题: 【{query}】")
print("-" * 50)

# 3. 计算向量编码
# GTE 模型会将这些文本转化成一串 768 维的数字
query_embedding = model.encode(query)
doc_embeddings = model.encode(documents)

# 4. 计算余弦相似度 (Cosine Similarity)
# 范围在 0 到 1 之间，越接近 1 表示语义越接近
cos_scores = util.cos_sim(query_embedding, doc_embeddings)[0]

# 5. 结果排序并直观展示
results = zip(documents, cos_scores)
sorted_results = sorted(results, key=lambda x: x[1], reverse=True)

print("语义检索结果（按匹配度排序）:")
for i, (doc, score) in enumerate(sorted_results):
    # 根据分数生成一个简单的进度条，视觉更直观
    bar_length = int(score * 20)
    bar = "█" * bar_length + "░" * (20 - bar_length)
    
    print(f"[{i+1}] 相似度: {score:.4f} | {bar} | 内容: {doc}")

Save, run—sample output:

5. Next level: FastAPI embedding service

In production, wrap inference as an API.

File: test2.py

from fastapi import FastAPI
from pydantic import BaseModel
from sentence_transformers import SentenceTransformer
import uvicorn

app = FastAPI()
model = SentenceTransformer('thenlper/gte-base-zh')

class TextRequest(BaseModel):
    input: list[str]

@app.post("/v1/embeddings")
async def get_embeddings(request: TextRequest):
    embeddings = model.encode(request.input).tolist()
    return {
        "object": "list",
        "data": [{"embedding": emb, "index": i} for i, emb in enumerate(embeddings)],
        "model": "gte-base-zh"
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)

Run:

python test2.py

This GPU box had no public IP—no browser check from outside.

On a host with a public IP, map 8000 or browse http://<public-ip>:8000 per your cloud rules.

6. Deployment tips

1. VRAM: GTE-Base-ZH weights ~400 MB; inference often 1–2 GB—very friendly.
2. Long text: default context ~512 tokens—chunk before embedding.
3. Quantization: onnxruntime quantization can speed CPU paths.
4. Batching: For index builds, model.encode(batch_size=32) vs one-by-one is often 10× faster.

7. FAQ

• Q: GTE-Small vs Base vs Large?
  • Small: fastest—edge / tiny VMs.
  • Base: best default balance.
  • Large: best recall, more VRAM—when every point of quality matters.
• Q: Why are all similarities > 0.9?
  • Embedding geometry—watch relative gaps (A closer than B) or L2-normalize (F.normalize) before cosine.

Hope this gets GTE running quickly—comment if you hit snags.

A prebuilt image is under review; I’ll link it here once approved.