Skip to content

Dzen Embedder

When an application needs vectors for semantic search, it can run the embedding model as a separate service. Dzen Embedder loads one SentenceTransformer model and accepts text over HTTP. Your application receives numerical vectors through an OpenAI-compatible embeddings API at POST /v1/embeddings and uses them in its search or RAG pipeline.

The project was built for Dzen Chat, but any application that can send HTTP requests can use it. The default model is deepvk/USER-bge-m3. It runs on CPU, NVIDIA GPUs through CUDA, or Apple Silicon through MPS.

A drop-in replacement for your existing OpenAI embeddings setup. Keep your client and application workflow: point the client to Embedder's /v1 base URL, set the configured model ID, and request encoding_format: float. This applies to clients using text inputs and supporting a custom endpoint and model. When switching embedding models, re-embed existing documents and use the new model for queries too. See API compatibility and limits.

Source code on GitHub Install Dzen Embedder

What the service provides

  • OpenAI-compatible embeddings API. Integrate with clients that support a custom endpoint and model ID, using the familiar embeddings request and response format with encoding_format: float. See API compatibility and limits.
  • One process with a loaded model and an HTTP API for embeddings.
  • Separate realtime and batch queues for user queries and background document processing.
  • Health endpoints at /live and /ready, and Prometheus metrics at /metrics.
  • An optional queue dashboard at /dashboard and built-in API examples at /docs.

Embedder does not store a search index, crawl websites, or generate answers with a language model. Document splitting, vector storage, and retrieval remain the application's responsibility.

Read about the benefits and trade-offs, then choose an installation method.