Resources · n8n integration

n8n NVIDIA Nemotron Embeddings nodeConfigure NVIDIA Nemotron Embeddings in n8n.

The n8n NVIDIA Nemotron Embeddings node turns text into the vectors a semantic search runs on. It is a sub-node, so it has no main input and never runs alone: it plugs into the Embeddings port of a root node. Five parameters cover the model, batching, text cleanup, vector size and timeout.

Try the NVIDIA Nemotron Embeddings node on n8nSee the 5 parameters

Verified Trustpilot reviews · AI, automation & growth agency

Why automate

What does the n8n NVIDIA Nemotron Embeddings node actually do?

It calls a NeMo Retriever embedding model, hosted by NVIDIA on build.nvidia.com or running as a self-hosted NIM container, and hands back the numeric representation of your text. In n8n terms it is a sub-node: a small node you attach to a bigger one rather than wire into the flow of items. The root node decides what gets embedded, this node decides which model does the embedding and how the requests are shaped.

Take a knowledge base first. Rows of policy text leave a Google Sheets node, reach a vector store in indexing mode, and the store asks this sub-node for a vector per row. You pick the model once in Model, either from the list or by typing the ID your NIM container serves, and the same choice applies to every row of the run.

Second case, the retrieval side of a support assistant. An AI Agent queries the store, the store embeds the question through this node, and the closest passages come back as context. The node sets input_type for you, passage while indexing and query while searching, so the same configuration covers both directions without a second node.

Third case, a private deployment. Point the credential at a NIM endpoint on your own infrastructure and nothing leaves it, while the workflow stays byte for byte identical to the cloud version. That is also where Dimensions earns its place: a Matryoshka model can return shorter vectors, which suits a store you want to keep small.

When would you reach for something else? When your text already sits behind a provider you have wired up for other reasons, an OpenAI credential for instance, a second key to manage buys you little. NVIDIA bills the calls it serves under its own terms, n8n adds nothing on top.

Two limits worth knowing before you build. The node is version 1 and exposes five parameters, no more: retries, custom headers and per-request overrides are not among them. And because it has no main input, you cannot test it on its own, you run the root node and read what comes out of the store. The n8n review covers the platform side, and the n8n training goes through a full indexing flow step by step.

Connect

Which credential does the node need?

  1. 01

    Create an NVIDIA build account

    Cloud access runs through build.nvidia.com, so start there. Sign in, open a Nemotron model in the catalogue and select Get API Key. The key belongs to the account, not to a single model, so one key covers every NeMo Retriever model you later select in the node. A self-hosted NIM container skips this step entirely.

  2. 02

    Add the NVIDIA Nemotron credential in n8n

    In the Credentials menu, create the NVIDIA Nemotron credential. It asks for a Base URL, the endpoint that follows the OpenAI spec, and an API Key. The default Base URL is https://integrate.api.nvidia.com/v1, which is what cloud models answer on. Paste the key you copied into API Key. Created once, the credential is reusable across every workflow.

  3. 03

    Or point the same credential at your NIM

    One credential covers both deployment modes. For a container of your own, replace the Base URL with your NIM endpoint and keep the /v1 path, as in http://localhost:8000/v1. If that container enforces authentication, paste its token into API Key. If it does not, leave the field blank, which is the one case where the key is optional.

Connections

Where does this sub-node plug in?

n8n splits AI work between a root node, which receives the workflow items, and sub-nodes attached to its ports, one type of sub-node per port. This node is a sub-node, and the port below is its only way into a workflow.

Output (what it plugs into)

  • Embeddingsai_embedding
01EmbeddingsThe ai_embedding port is what this node feeds. A root node that stores or searches vectors exposes it, and until the connection exists the node has nothing to run on. The node can only attach where that port is offered, never on a chat model or memory port.

The ai_embedding port is what this node feeds. A root node that stores or searches vectors exposes it, and until the connection exists the node has nothing to run on. The node can only attach where that port is offered, never on a chat model or memory port.

Key parameters

  • Required: with no main input of its own, the node only executes when a root node asks it for vectors.
Use cases
a vector store in indexing mode calls the port while filling itself, then calls it again on every question it has to answer.
Parameters

What do the 5 parameters change?

The NVIDIA Nemotron Embeddings node has 5 parameters. For each one: the node as you configure it in n8n, what the parameter changes, and our field notes.

01

Model

model

What you see in n8n

Notes & use cases

This is the only required parameter: it names the NeMo Retriever embedding model that produces the vectors. It is a resource locator, so you either pick from the list loaded through the credential or type an ID directly, which is how a self-hosted NIM model gets selected.

Key parameters

  • Model: required. The node sets input_type on its own, passage when indexing and query when searching, so nothing to switch between the two.
Use cases
keep one model for a given store, since vectors written by one model are only comparable with vectors from that same model.
02

Batch Size

options.batchSize

What you see in n8n

Notes & use cases

Grouping matters once a run stops being a handful of rows. This option sets the maximum number of documents packed into each request to the model, which is what decides whether a long import goes out in a few calls or in hundreds.

Key parameters

  • Batch Size: a number, the ceiling per request. Lower it when a large import starts failing, raise it when an index run drags.
Use cases
indexing a document base overnight, where the batch size is the difference between a job that finishes and one that times out halfway.
03

Strip New Lines

options.stripNewLines

What you see in n8n

Notes & use cases

Raw text rarely arrives clean. Turned on, this switch removes the line breaks from the input before it is sent, which is worth doing on content that carries formatting it never asked for.

Key parameters

  • Strip New Lines: a checkbox, on or off. Apply the same setting when indexing and when searching so both sides see text shaped the same way.
Use cases
a PDF converted to text, or an email body pulled from Gmail, where every wrapped line would otherwise land in the vector.
04

Dimensions

options.dimensions

What you see in n8n

Notes & use cases

Vector length is negotiable on some models. Setting a number here asks for output embeddings of that size, and leaving it unset keeps whatever the model returns by default.

Key parameters

  • Dimensions: a number, and only models with dynamic Matryoshka embeddings honour it. On any other model, leave it alone.
Use cases
trimming a store that has grown past what you want to host, once you accept that the whole index has to be rebuilt at the new size.
05

Timeout

options.timeout

What you see in n8n

Notes & use cases

Requests that never answer are worse than requests that fail. This value caps how long a single call is allowed to take, expressed in seconds, and -1 removes the cap altogether.

Key parameters

  • Timeout: a number of seconds. A NIM that has just started serving is the usual reason to raise it.
Use cases
a scheduled reindex you would rather see stop and report than sit open, or the opposite, a slow container you allow to take its time.
Need help

Need help automating NVIDIA Nemotron Embeddings with n8n?

A person reads every message.

FAQ

Questions people ask about this node

01Is the NVIDIA Nemotron Embeddings node included in n8n?
Yes. It ships inside the n8n AI package, so it is there on n8n Cloud and on a self-hosted instance alike, with nothing to install and no extra charge on the n8n side. The Community Edition under the Sustainable Use licence includes it too, and a workflow built on one behaves the same on the other. What does have a cost is the model itself: NVIDIA bills the calls it serves under its own terms. A NIM container you run yourself changes that picture, since the infrastructure is then yours.
02What do you need for it to work?
A credential and a root node to attach it to. The credential is the NVIDIA Nemotron one, which asks for a Base URL and an API Key. The key is required for build.nvidia.com and optional for a self-hosted NIM that does not enforce authentication, in which case the field stays empty. The Base URL defaults to the NVIDIA cloud endpoint and gets replaced by your own when you self-host. Beyond that, the node needs an Embeddings port to hang off, because it has no main input and never executes on its own.
03What are the limits of the node?
It is version 1 with five parameters, and that is the whole surface: the model, plus batch size, new line stripping, dimensions and timeout. Anything else you might want per request is not exposed. Dimensions only works on models with dynamic Matryoshka embeddings, so setting it elsewhere buys nothing. And because it is a sub-node, it cannot be executed or tested by itself, which makes the root node your only debugging view. Older workflows may still show an earlier version with fewer options.
04What has to be connected for the n8n NVIDIA Nemotron Embeddings node to run?
One root node, connected on the Embeddings port. That port carries the ai_embedding type, so the node attaches to a root node that stores or searches vectors, and to nothing else. Start on the indexing side: send a small batch of documents through, confirm that the store fills up, then reuse the exact same node and the exact same model on the search side. Keeping one model per store matters, because vectors produced by two different models cannot be compared against each other.
05n8n or Make for this kind of embedding workflow?
It depends on where the data has to live. Make is hosted by Make with no self-hosting option and charges per operation, which is comfortable when the volume stays predictable and nothing sensitive is involved. n8n can be self-hosted, and paired with a NIM container on the same infrastructure it lets an indexing pipeline run without the text leaving your network. Volume pushes the same way: an import that embeds thousands of rows counts operations quickly on a per-operation plan.
Hack'celeration Lab

Get our weekly integration tips.

No spam. Unsubscribe anytime.

Same platform

Other docs you might need on this stack.

Browse all integrations