Remote embeddings (OpenAI-compatible)

Semango’s openai provider uses the OpenAI Embeddings API and supports OpenAI‑compatible endpoints via base_url.

Minimal config

embedding:
  provider: openai
  model: text-embedding-3-large
  api_key_env: OPENAI_API_KEY

OpenAI-compatible providers

If you use a compatible provider (self-hosted or third‑party), set base_url or base_url_env:

embedding:
  provider: openai
  model: text-embedding-3-small
  api_key_env: MY_EMBEDDING_API_KEY
  base_url_env: MY_OPENAI_BASE_URL

Semango will pass requests to base_url using the OpenAI-compatible schema.

Known model dimensions

Semango has a small list of known embedding dimensions (used for validation):

• text-embedding-3-large → 3072
• text-embedding-3-small → 1536
• text-embedding-ada-002 → 1536
• text-embedding-nomic-embed-text-v1.5 → 768

Manual Dimension Override (Truncation)

You can specify a custom dimension using the dim parameter.

embedding:
  provider: openai
  model: text-embedding-3-large
  dim: 1024  # Truncate large embeddings to 1024

For OpenAI text-embedding-3-* models, Semango passes this value directly to the API's dimensions parameter, which is more efficient than manual truncation. For other models (including self-hosted compatibles), Semango truncates the vector returned by the endpoint to the first dim elements.

IMPORTANT

Matryoshka Support: Truncation is most effective with models trained via Matryoshka Representation Learning (e.g., OpenAI's v3 models). For older models like text-embedding-ada-002, truncating the vector will likely lead to poor search performance. Check sbert.net for more information on compatible models.