Embedding Models

Overview and configuration of supported embedding models

Embeddings turn your documents into searchable vectors. Which model you use depends mostly on your deployment path.

Recommended: built-in offline model (Desktop app)

The Desktop app ships with a built-in embedding model — the best choice for most teams:

  • Model: e5-small (int8-quantized), optimized for German
  • Offline: runs entirely locally, with no external service and no API key
  • No setup: no Ollama, no GPU, no configuration — ready out of the box

If you use the Desktop app, that's all you need. Read on only if you want to force a different model.

Advanced: your own embedding provider

If you want full control — typical on the Docker path — you configure your embedding provider yourself. They all use the same OpenAI-compatible interface.

Local / self-hosted (Ollama & co.)

  • Embeddings stay inside your own environment
  • The model runs outside the Outpost process (e.g. as an ollama service)
  • The Outpost uses provider: "openai" and the service base_url
  • Common choice in the Docker setup: bge-m3 (1024 dim, multilingual) — an option, not a mandate

Cloud providers

OpenAI

Cloud-based embeddings via the OpenAI API

Azure OpenAI

OpenAI models via Microsoft Azure

Nebius

Multilingual embeddings optimized for European languages

Basic configuration

embedding_model:
  provider: "openai"  # or azure, nebius
  # Provider-specific parameters…

All providers support these common parameters:

ParameterTypeDefaultDescription
providerstring–Provider name
rpminteger3000Requests per minute
tpminteger1000000Tokens per minute

Selection guide

  • Simplest path: the Desktop app's built-in offline model — German-optimized, zero setup
  • Local / self-hosted: Ollama or another OpenAI-compatible service — embeddings stay with you
  • Cloud (OpenAI, Azure, Nebius): no local hardware, pay-per-use, quick to start

Changing the model means reindexing

Changing the embedding model requires a complete reindex of all documents. Choose deliberately before your initial setup.

Comparing embedding quality

Test different models with your own documents:

  • Language: German and English search
  • Domain: domain-specific vs. general content
  • Length: short vs. long text passages
  • Cost: API cost vs. hardware investment

Approach: upload a small document collection, run typical search queries, evaluate the relevance of results and response times.

Migrating between models

  1. Preparation: back up the current configuration, test the new model separately, plan downtime
  2. Stop the Outpost
  3. Activate the new embedding_model configuration
  4. Restart the Outpost
  5. Wait for the complete reindex

Support & troubleshooting

  • API limits: adjust rpm/tpm
  • Authentication: check API keys
  • Performance: check service latency and embedding batch size

For questions about embedding configuration: 📧 Enterprise Support: enterprise@meingpt.com

Was this page helpful?