mxen/dzen-embedder repository overview

⁠Dzen Embedder

A single-process, OpenAI-compatible text embedding service with one shared SentenceTransformer model and separate realtime/batch queues.

Source and Compose files: dzenplatform/embedder⁠. License: BSD-3-Clause. Model weights have their own licenses and are downloaded separately.

⁠Image variants

Moving aliases: cpu, cu126, and latest (CPU). Docker selects the CPU architecture automatically. Use a versioned tag or digest in production. These are project-maintained images; they are not part of Docker’s curated Docker Official Images program.

⁠Docker Compose

From the source checkout:

# CPU
docker compose up -d

# NVIDIA GPU: use this instead of the CPU command
docker compose -f compose.yaml -f compose.cuda.yaml up -d

Compose downloads the default deepvk/USER-bge-m3 model to a persistent volume before starting the service. The API is available at http://127.0.0.1:8091; readiness is /ready, documentation is /docs, and embeddings use POST /v1/embeddings.

Set EMBEDDER_TAG, EMBEDDER_PORT, MODEL_ID, or GPU_DEVICE_ID as needed. GPU 0 is selected by default.

NVIDIA requires a compatible host driver and NVIDIA Container Toolkit⁠. The image tag alone does not grant GPU access. Jetson and CUDA arm64 are not included; newer GPU generations may require a newer CUDA/PyTorch variant. GPU inference for this first image still requires validation on the target hardware.

Apple MPS requires native macOS and is not available inside this Linux image. Use CPU in Docker on a Mac, or the native MPS setup documented in the source repository.