Griptape Embedding Driver: OpenAI Compatible

The most versatile embedding driver in the pack: Griptape Embedding Driver: OpenAI Compatible produces embeddings from any service that speaks OpenAI's embeddings API - OpenAI proper, a proxy, a local server with an OpenAI-compatible layer, or a hosted clone. You point base_url at whatever you want, and the same node keeps working.

This is the driver you reach for when your provider isn't in the pack's named list, or when you route through a gateway. If you're on plain OpenAI, the dedicated OpenAI driver is simpler; this one earns its keep in heterogeneous setups.

How it works

Configuration node, pack-standard: it builds an EMBEDDING_DRIVER that downstream RAG/vector-store nodes use to convert text into vectors, with your endpoint doing the inference. The LM Studio driver in this pack is literally a subclass of this one - that's how much "compatible" is doing the heavy lifting here.

The inputs:

• embedding_model - the model ID your endpoint serves, default text-embedding-3-small. Change it to match what your provider actually offers.
• base_url - https://api.openai.com/v1 by default; repoint at your proxy or provider.
• api_key_env_var - env-var name for the key, default OPENAI_API_KEY. Some local endpoints accept any string; others need a real key - match your provider.

Installing

Ships in the ComfyUI Griptape Nodes pack:

• ComfyUI Manager: search "Griptape" → install ComfyUI-Griptape.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/griptape-ai/ComfyUI-Griptape, then restart.

Pack dependencies: griptape[all], openai, python-dotenv, plus git-hosted extensions. The torch caveat applies pack-wide (reinstall torch with the CUDA index if Griptape's install conflicts with ComfyUI's build).

Gotchas

"Compatible" is a spectrum, not a guarantee: a provider can speak the embeddings API but serve a different model list, require a different auth header, or return vectors with different dimensions. When a call fails, test the endpoint directly with curl before blaming the node. The dimension point matters more here than anywhere - because you can freely swap endpoints, it's easy to end up with a vector store holding mixed-dimension vectors. If searches start returning garbage after a provider swap, that's why; rebuild the index.