Overview
- Full name
docker.io/semitechnologies/reranker-transformers- Registry
- docker.io
- Namespace
- semitechnologies
- Repository
- reranker-transformers
- Cached tags
- 28
- Last synced
- 07/06/2026, 02:25 PM
Introduction
Weaviate 生态 Cross-Encoder 重排序模型服务镜像，对检索候选文档二次打分，提升 RAG 与向量搜索的 Top-K 精度。
Details
semitechnologies/reranker-transformers 加载 HuggingFace Cross-Encoder 模型，接收 query+passage 对返回相关性分数，通常部署在 Weaviate 或自研 RAG 管道下游。典型场景是向量召回后精排以改善答案质量。与 ColBERT 相比延迟更高但实现简单；GPU 可选加速，CPU 模式适合中小 QPS。Kubernetes 中作为无状态 Deployment 水平扩展，模型缓存需足够 memory。
快速启动
docker run -d --name reranker \
-p 8080:8080 \
docker.io/semitechnologies/reranker-transformers:latest
推荐实践
docker run -d --name reranker \
--restart unless-stopped \
-p 8080:8080 \
-e RERANKER_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2 \
-e ENABLE_CUDA=0 \
docker.io/semitechnologies/reranker-transformers:2.0.0
核心参数说明
-p 8080:8080— 重排序 HTTP API 端口-e RERANKER_MODEL— HuggingFace 模型 ID-e ENABLE_CUDA— GPU 推理开关（0=CPU）--restart unless-stopped— RAG 链路持续可用- 固定 tag — 与 Weaviate 模块版本配套
- 资源 — 模型加载需预留 1Gi+ 内存
Kubernetes
apiVersion: apps/v1
kind: Deployment
metadata:
name: reranker-transformers
spec:
replicas: 2
selector:
matchLabels:
app: reranker-transformers
template:
metadata:
labels:
app: reranker-transformers
spec:
containers:
- name: reranker
image: docker.io/semitechnologies/reranker-transformers:2.0.0
ports:
- containerPort: 8080
env:
- name: RERANKER_MODEL
value: cross-encoder/ms-marco-MiniLM-L-6-v2
resources:
requests:
memory: 2Gi
---
apiVersion: v1
kind: Service
metadata:
name: reranker-transformers
spec:
selector:
app: reranker-transformers
ports:
- port: 8080
targetPort: 8080
Weaviate 或检索服务经 ClusterIP 调用；GPU 节点可设 nodeSelector 与 ENABLE_CUDA=1，按 QPS HPA 扩副本。