Aller au contenu

Embedding / RAG

Embedding models convert conversation text into vectors used by semantic conversation search. Configure them under Settings → Conversation Search.

Providers and presets

Available presets currently include:

  • OpenAI: text-embedding-3-small, text-embedding-3-large, text-embedding-ada-002
  • Mistral: mistral-embed
  • Voyage AI: voyage-3-large, voyage-3-lite, voyage-code-3
  • SiliconFlow: BAAI/bge-m3, BAAI/bge-large-en-v1.5
  • OpenRouter: OpenAI embedding model routes
  • Ollama, a local embedding model, or a custom endpoint

Remote embeddings use the credentials and base URL configured for the selected provider. Text sent for embedding therefore leaves the device for that provider. Local embeddings remain on-device.

RAG controls

  • Context range: 4–32 conversation steps, in steps of 4
  • Result count: 5–30, in steps of 5
  • Similarity threshold: 0–1 (default 0.5)

Changing the embedding model may require existing indexed content to be embedded again before results are complete. See Conversation Search and Privacy & Security.