weaviate-sbert-vectorizer

CPU-only Sentence Transformers vectorizer (embedding) service for Weaviate.

Vectorization / Embedding is the process by which human-understandable data (ie. text) is converted into machine-readable numerical representations (vectors) for use in AI systems. Vectors are stored in specialized vector databases that are optimized for rapid and nuanced information retrieval.

This service is designed for CPU vectorization on a single core using the ONNX Runtime. No GPU is required.

Built images contain only the minimum dependencies required, and are thus significantly smaller than those provided by Weaviate (by ~7GB).

Usage

The images of this service are intended to be used with Weaviate's text2vec-transformers module:

services:
  weaviate:
    environment:
      ENABLE_MODULES: text2vec-transformers
      TRANSFORMERS_INFERENCE_API: http://wstv:8080
  wstv:
    image: ghcr.io/metabronx/weaviate-sbert-vectorizer:all-MiniLM-L6-v2_quint8_avx2

Configuration

The embedding model is configurable via the MODEL_NAME and FILE_PATH environment variables, which correspond:

SentenceTransformer(
    MODEL_NAME,
    model_kwargs={"file_name": FILE_PATH},
    backend="onnx",
)

The default model is the 8-bit-quantized AVX2-optimized (quint8_avx2) variant of sentence-transformers/all-MiniLM-L6-v2.