BYOK Providers

Choosing a Provider

All providers produce embedding vectors for semantic search. The key difference is dimension support — your account locks to one dimension on first key save.

Recommended: 1024 dimensions — supported by OpenAI, Gemini, Cohere, and Voyage.

OpenAI

• Model: text-embedding-3-small (1536d default, supports custom 256-3072)
• LLM: Yes (gpt-4)
• Supports baseUrl for OpenRouter, Together, Fireworks

Google Gemini

• Model: gemini-embedding-001 (3072d default, supports custom 256-3072)
• LLM: Yes (gemini-2.5-pro)
• Free tier: unlimited embedding calls

Cohere

• Model: embed-v4.0 (1024d default, supports 256-1024)
• LLM: No (embedding only)
• Free tier: 1,000 calls/month
• Supports query/document input types for better search quality

Voyage AI

• Model: voyage-3 (1024d, supports custom)
• LLM: No (embedding only)
• Free tier: 200M tokens, 3 RPM limit

Azure OpenAI

• Requires base URL: https://your-resource.openai.azure.com
• Key format: API key + base URL in the save form
• Same models as OpenAI

Mistral

• Model: mistral-embed (256d fixed, cannot be changed)
• LLM: Yes (mistral-large-latest)
• Only compatible with accounts locked at 256d

Ollama (Local)

• Model: nomic-embed-text (768d)
• API key = endpoint URL (e.g., http://localhost:11434/v1)
• Free, self-hosted

Qwen (Self-Hosted)

• Model: Qwen3-Embedding-8B (4096d)
• API key = TEI endpoint URL
• Free, self-hosted via HuggingFace TEI