Griptape Embedding Driver: Amazon SageMaker Jumpstart

This is the "bring your own model" embedding driver. Where the other embedding nodes point at a managed API, Griptape Embedding Driver: Amazon SageMaker Jumpstart points at a SageMaker endpoint that you deployed - typically an embedding model launched through Jumpstart (Sentence Transformers, BGE, that crowd) running on your own AWS infrastructure.

It's the most work of any embedding driver in the pack, and it's only worth it if you need it: model choice, endpoint sizing, and billing all live on your side of the fence. For a quick RAG experiment it's overkill - Ollama or OpenAI get you there in minutes. For a locked-down, self-hosted embedding service that stays inside your VPC, it's the right tool.

How it works

Configuration node, same shape as the other embedding drivers: it builds an EMBEDDING_DRIVER object that downstream nodes (vector store, RAG) use to turn text into vectors. The endpoint does the actual inference.

The inputs:

• embedding_model - freeform string, default empty. This is the model you deployed at the endpoint; it's informational/used in the driver config, and it's up to you to keep it in sync with what the endpoint actually runs.
• endpoint - the SageMaker endpoint name (the default placeholder is jumpstart-dft-...). This is the load-bearing input: it has to match a real, running endpoint in your account.
• The AWS trio - aws_access_key_id_env_var, aws_secret_access_key_env_var, aws_default_region_env_var - env-var names, not literal keys, same as every AWS driver in the pack.

The node description on the brief is honest but sloppy ("OpenAI Compatible Embedding Driver") - don't let that throw you. The source shows a SageMaker embedding driver backed by your endpoint.

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 pack-wide torch caveat applies (reinstall torch with the CUDA index if Griptape's install conflicts with ComfyUI's - README troubleshooting).

Gotchas

The endpoint must exist and be in service, or you'll get a "could not connect to endpoint" error that looks like a credentials problem. Deploying a Jumpstart embedding model takes a few minutes and starts billing the moment it's live - people forget the endpoint is running and eat a monthly charge for a model they used once. Also keep the region consistent: the endpoint and the driver's region env var have to agree. And don't skip setting the embedding model name to match your deployment, even though the default is blank - a mismatch is a debugging rabbit hole.