Embedding dimensions are truncated to 256 in some cases

[dflatline]

Confirm this is a Node library issue and not an underlying OpenAI API issue

• This is an issue with the Node library

Describe the bug

With some API endpoints, this package truncates the embedding dimension to 256. For models with a larger embedding dimension than 256, this is a massive accuracy loss.

So far, ollama and LM Studio openai-compatible endpoints suffer from this issue. llama-server does not.

This issue impacts RooCode (see RooCodeInc/Roo-Code#4462).

To Reproduce

Download embedding an model with more than 256 dimensions in ollama and LM Studio. For example, snowflake-arctic-embed-large-v2 has 1024 dimensions.
ollama pull Definity/snowflake-arctic-embed-l-v2.0-q8_0

Then, run the attached node script, and notice that the package truncates the expected dimensions of 1024 to 256. If the same model from LM Studio is served from llama-server, there is no problem:
llama-server -m ~/.cache/lm-studio/models/Casual-Autopsy/snowflake-arctic-embed-l-v2.0-gguf/snowflake-arctic-embed-l-v2.0-q6_k.gguf --embeddings -c 4096 -ngl 99

package.json

test-openai-embeddings.js.txt

Code snippets

OS

macOS

Node version

Node v23.11.0

Library version

openai 4.104.0

[sokdara420]

Bug