I send this request:

curl -X POST http://localhost:8000/v1/embeddings \ -H "Authorization: Bearer your-secret-token" \ -H "Content-Type: application/json" \ -d '{"input": "Hello world", "model": "nomic-embed-text", "dimensions":1}'

But I still got 768 dimensions. Can you check @copilot, why?

I send this request:

`curl -X POST http://localhost:8000/v1/embeddings
...

The nomic-embed-text model always outputs 768-dimensional embeddings - it doesn't support the dimensions parameter. Ollama's API doesn't officially support dynamic dimension sizing for most models.

The implementation correctly passes the parameter through to Ollama, but it has no effect because:

1. nomic-embed-text has a fixed 768-dimension output
2. Ollama's embedding API doesn't honor the dimensions parameter for standard models

Only models with Matryoshka Representation Learning (MRL) capability can support dimension reduction, and even then, Ollama doesn't expose this through the API - you'd need to manually truncate the embeddings after retrieval.

If you need smaller embeddings, consider:

• Using a model with naturally smaller dimensions (e.g., all-minilm with 384 dimensions)
• Post-processing embeddings to truncate them manually
• Using a different embedding service that supports dimension control

The parameter implementation is correct for API compatibility with OpenAI, but it won't affect Ollama's output with current models.

@copilot based on this doc, ollama supports: https://docs.ollama.com/api/openai-compatibility

@copilot based on this doc, ollama supports: https://docs.ollama.com/api/openai-compatibility

You're absolutely right! I apologize for the incorrect information. Ollama does support the dimensions parameter as of version v0.11.11 (released in 2024).

The implementation is correct and will work with:

• Ollama v0.11.11 or later
• Models that support dynamic dimensions (e.g., Qwen3 Embedding with MRL support)

However, nomic-embed-text has fixed 768 dimensions and won't change its output based on the dimensions parameter. To test the feature, use a model with MRL (Matryoshka Representation Learning) support like Qwen3 Embedding.

I've updated the documentation in commit 98dfc94 to correctly reflect that Ollama supports this parameter (model-dependent).