Embedding Providers

Embeddings turn your text into vectors so OpenLoomi can search it semantically. OpenLoomi can generate embeddings either through a cloud provider or a local, on-device model. Configure this in Embedding settings (surfaced within AI Settings).


Overview

Embeddings power Memory semantic search, the knowledge-base RAG, and the semantic groups in Global Search. Every stored item is embedded once and compared by cosine similarity at query time. The provider you pick determines where that embedding computation happens.


Cloud provider

The cloud path sends text to a hosted embedding API.

  • Default base URL: https://openrouter.ai/api/v1
  • Default model: text-embedding-3-small

Provide an API key for your provider (OpenAI or OpenRouter-compatible). This path requires network access and an embedding API key.


Local provider

The local path runs an embedding model on your machine — no data leaves your device. Built-in options:

ModelBest forDimensions
Xenova/all-MiniLM-L6-v2English384
Xenova/bge-large-zh-v1.5Chinese1024

You can also supply a custom Hugging Face model ID or local path. A local-files-only switch lets you disable model downloads once a model is cached, so the app never reaches the network for embeddings.


Switching providers

You can switch between cloud and local, and between models, from the embedding settings. Keep in mind:

  • Different models produce different vector dimensions (e.g. 384 vs 1024 vs 1536). Existing vectors were computed with the old model, so switching means new content is embedded with the new model.
  • To make search consistent after a switch, existing content needs to be re-embedded with the new model. Until that happens, retrieval quality can dip because queries and stored vectors were produced by different models.
  • The settings remember the last model you used per provider type, so toggling back restores your previous choice.

Troubleshooting

  • "I switched models but vectors still look like the old dimension." Stored vectors keep their original dimension until re-embedded. New items use the new model immediately; existing items need a re-embed pass.
  • Retrieval quality dropped after switching. Mismatched query/stored embeddings degrade similarity. Let re-embedding complete, or switch back to the model your existing content was embedded with.
  • Local model won't load. If local-files-only is on, the model must already be cached. Turn it off once to allow the initial download, then re-enable it.
  • Cloud embeddings fail. Confirm the base URL and API key. Semantic search throws when no embedding provider is configured.