With LM Studio serving nomic-embed-text-v1.5:

• curl request without encoding_format → returns 768-dim float array ✅
• OpenAI SDK request without explicit encoding_format → LM Studio returns 192-dim array ❌

TL;DR

To fix the issue, add encoding_format: "float" to the embeddings request to ensure consistent behavior with local embedding providers.

Guidance

• Verify that the local embedding provider (e.g., LM Studio) is configured to handle float encoding correctly.
• Check the OpenAI SDK version to confirm it is v6.34.0, which changed the default encoding_format to base64.
• Update the memory-lancedb/index.js file to include encoding_format: "float" in the embeddings request, as shown in the provided fix.
• Test the fix by running a curl request and an OpenAI SDK request without explicit encoding_format to ensure both return the expected 768-dim float array.

Example

The fix involves updating the embed function in memory-lancedb/index.js to include the encoding_format parameter:

async embed(text) {
    const params = {
      model: this.model,
      input: text,
      encoding_format: "float"
    };
    // ...
}

Notes

This fix assumes that the local embedding provider (LM Studio) is correctly handling float encoding. If issues persist, verify the provider's configuration and compatibility with the OpenAI SDK.

Recommendation

Apply the workaround by adding encoding_format: "float" to the embeddings request, as this ensures consistent behavior with local embedding providers and resolves the issue caused by the OpenAI SDK's default encoding_format change.