May a use models from OpenRouter?

[kaanoz1]

First off, I am not certain that declared ai models in configurator page is locally are locally running or fetching outputs from somewhere else. May anybody illuminate me on that topic?

And my other question is can I use models from open router? While my configuration as follows:

I set the OPENAI_API_KEY as my environment variable with the key that OpenRouter provided. Can I use models in that way? If yes. May you help you about the error I got while I get my errors validate at the last stage of the configuration phase?

My final configuration:

[vendor]
simple_model = "openai"
complex_model = "openai"
embeddings_model = "openai"

[aws]
simple_model = "anthropic.claude-3-haiku-20240307-v1:0"
complex_model = "anthropic.claude-3-5-sonnet-20240620-v1:0"
embeddings_model = "amazon.titan-embed-text-v1"
region_name = "us-east-1"

[openai]
simple_model = "gpt-4o-mini"
complex_model = "gpt-oss-20b:free"
embeddings_model = "text-embedding-3-small"
base_url = "https://openrouter.ai/api/v1/"

[ollama]
simple_model = "llama3.2"
complex_model = "llama3.1:70b"
embeddings_model = "llama3.2"
base_url = "http://localhost:11434"

[tracing]
project = "rai"

[tracing.langfuse]
use_langfuse = false
host = "http://localhost:3000"

[tracing.langsmith]
use_langsmith = false
host = "https://api.smith.langchain.com"

[asr]
recording_device_name = "default"
transcription_model = "LocalWhisper"
language = "en"
vad_model = "SileroVAD"
silence_grace_period = 0.3
vad_threshold = 0.3
use_wake_word = false
wake_word_model = ""
wake_word_threshold = 0.5
wake_word_model_name = ""
transcription_model_name = "tiny"

[tts]
vendor = "ElevenLabs"
voice = ""
speaker_device_name = "default"

The error I got:

I did not get why I am getting error ONLY on embeddings model while complex model works.

[Tutku055]

I am fairly certain this is a routing issue within the application's config logic.

The error message points to platform.openai.com, which proves the request is hitting the official OpenAI API directly, bypassing your OpenRouter base_url.

Since you provided an OpenRouter key, the official OpenAI endpoint rejects it as invalid (401).

The fact that 'Simple' and 'Complex' models passed the test confirms your key and base_url are correct. The application likely fails to apply the base_url specifically for the Embeddings client.

Fix: You might need to verify if there is a specific embeddings_base_url setting you missed, or this might be a bug in the software where it hardcodes the endpoint for OpenAI embeddings."

[Juliaj]

Give a try of this PR #756. Let us know whether it resolves the issue with embedding model.

[Tutku055]

Hi @Juliaj ,

Thanks for the update and the commit! I've reviewed the latest changes in model_initialization.py and configurator.py.

Since I don't have a local OpenRouter setup to verify runtime behavior, I relied on static code analysis. Explicitly passing the base_url to the OpenAIEmbeddings constructor is exactly the correct fix for the routing issue. Also, seeing the green check ✅ on the CI confirms the build is stable.

The logic looks perfect. Good to merge! 🚀