Resources · n8n integration

n8n Milvus Vector Store nodeConfigure Milvus Vector Store in n8n.

Milvus holds the vectors, this node decides what happens to them. Two head parameters sit in the panel, and one of them switches the n8n Milvus Vector Store node between four operation modes: ingestion, ranked search, and retrieval for a chain or an agent. For builders wiring a RAG pipeline.

Try the Milvus Vector Store node on n8nSee the 2 parameters

Verified Trustpilot reviews · AI, automation & growth agency

Why automate

What does the n8n Milvus Vector Store node actually do?

The n8n Milvus Vector Store node connects a workflow to a Milvus database used as a vector store, meaning a database that keeps mathematical representations of your content instead of raw text. It writes documents into a collection, reads ranked documents back for a query, or exposes the collection to AI nodes so a chain or an agent can search it while answering. One node, four modes, selected in the panel.

Start with ingestion. In Insert Documents mode the node takes the items reaching its main input and writes them into the collection you picked. That is the pattern behind a knowledge base rebuilt on a schedule: fetch the source files, split them, embed them, push them into Milvus. The node options carry a Clear Collection switch for that exact job, so a rebuild starts from an empty collection instead of stacking a second copy of every paragraph on top of the first.

Then reading. Get Many mode turns a Prompt into a similarity search and hands back the best matches, capped by Limit. This is the mode to reach for when the workflow does the reasoning itself and Milvus is only the lookup step: a support ticket arrives, the node pulls the closest procedure pages, a model drafts the reply, Slack posts it for review. Rerank Results reorders what comes back before the model ever sees it, which matters when the top hit and the fourth hit are close.

The last two modes hand the collection to something else rather than returning rows. Retrieve Documents (As Tool for AI Agent) plugs the store straight into the tools connector of an AI Agent, which then queries it on its own when a question needs your data. Its Name and Description are what the model reads to decide whether to call the tool at all, so a vague description is the usual reason an agent ignores a perfectly good knowledge base. Same logic if you build the tool side with AI Agent Tool.

When should you pick something else? If the workflow is a straight question and answer flow with no decisions to make, Retrieve Documents (As Vector Store for Chain/Tool) feeding a Vector Store Retriever and a Question and Answer Chain is simpler than an agent, and cheaper to debug. Embeddings stay a separate sub-node in every case, so the model provider you pick there, OpenAI or Anthropic among others, bills its own API calls. Limits worth knowing: the node is at version 1, the collection has to exist on the Milvus side before the node can touch it, and a required field left empty stops the execution. If the modes do not cover what you need, the n8n training pages go deeper on wiring the rest of the pipeline.

Connect

What do you connect before the node runs?

  1. 01

    Create the Milvus credential once

    Open the Credentials menu in n8n and create the Milvus credential: Base URL (defaults to http://localhost:19530), Username (defaults to root) and Password (defaults to Milvus). Change those defaults before production. Credentials live outside the workflow, so the same entry is reused everywhere a rotated secret updates in one place.

  2. 02

    Select it on the node

    Back in the node panel, pick that credential in the credential field at the top. There is no Authentication selector on this node, so nothing else to choose: one credential, and the four values of Operation Mode all run through it. Insert and Get Many hit the same database with the same access, which is why a write that fails usually fails for the credential, not for the mode.

  3. 03

    Point it at a collection and run one item

    Fill Milvus Collection, either by picking it from the list or by typing the name, then run the node on a single item before wiring anything downstream. Get Many with a short Prompt is the quickest smoke test: documents come back, the connection is good. Nothing comes back and you know the problem is the collection or the credential, not the rest of the workflow.

Parameters

Which parameters do you set in the panel?

The Milvus Vector Store node has 2 parameters. For each one: the node as you configure it in n8n, what the parameter changes, and our field notes.

01

Operation Mode

mode

What you see in n8n

Notes & use cases

This selector decides everything else in the panel. Pick the mode first, then fill the fields it reveals.

Key parameters

  • Operation Mode: insert writes documents into the collection, load returns ranked documents for a query, retrieve exposes the store to a chain or tool, retrieve-as-tool exposes it to an agent.
  • Prompt: required in Get Many, the search text ranked against the collection.
  • Limit: how many top results come back.
  • Name and Description: required in tool mode, what the model reads to decide whether to call the store.
  • Include Metadata and Rerank Results: return the document metadata, reorder the hits before returning them.
Use cases
one collection, an insert branch rebuilding it nightly and a tool branch answering questions from it.
02

Milvus Collection

milvusCollection

What you see in n8n

Notes & use cases

The collection the node reads from or writes to. It is required in all four modes, so it is the one field you always fill.

Key parameters

  • Milvus Collection: a resource locator, so either pick the collection from the list loaded through the credential, or type its name, typically {{ $json.collection }} when a previous node decides which one to hit.
Use cases
routing by tenant. One workflow, the collection name arriving with the item, and each customer keeps their documents in their own collection instead of sharing a filtered one.
Need help

Need help automating Milvus Vector Store with n8n?

A person reads every message.

FAQ

Questions builders ask next

01Is the Milvus Vector Store node included in n8n?
Yes. It ships inside the AI package of n8n, so there is nothing to install and nothing extra to pay on the n8n side. It behaves the same on n8n Cloud, the hosted offer run by n8n, and on a self-hosted instance installed with Docker or npm under the Community Edition and its Sustainable Use license. A workflow built on one runs on the other. What you do pay for sits outside n8n: your Milvus database, and the model provider behind the embeddings sub-node, which bills its own API calls under its own terms.
02What do you need to authenticate for the n8n Milvus Vector Store node?
One thing on the node itself: a Milvus credential, created once in the Credentials menu of n8n and then reused by every workflow that talks to the same database. The node has no Authentication selector, so there is no method to choose, and the same credential serves all four operation modes. Two things are worth separating from that. The embeddings sub-node you connect carries its own provider credential, which is not the Milvus one. And the collection itself has to exist on the Milvus side: a valid credential pointed at a missing collection still fails.
03What are the limits of this node?
The node is at version 1, and an older workflow may still show an earlier version with fewer options, so what you see in your panel can differ from a freshly added node. The scope is the four operation modes and nothing beyond them: writing documents, ranked reading, and exposing the collection to a chain or an agent. Administration of Milvus itself is out of scope, and the collection has to be created on the database side first. Required fields are strict, an empty Milvus Collection or an empty Prompt stops the execution on the first item.
04What do you connect to it at minimum, and where should you start?
At minimum, an embeddings sub-node, because the node stores and searches vectors rather than text, and something that supplies the documents on the insert side. A text splitter and a data loader are the usual pair feeding it. For a first build, stay on the regular node patterns: Insert Documents in one workflow to fill the collection, Get Many in another to read it back. Move to Retrieve Documents (As Tool for AI Agent) once the searches return what you expect, so a disappointing answer points at the agent rather than at the indexing.
05n8n or Make for a vector store workflow?
It depends on where the data has to live. Make is hosted only and billed per operation, which is predictable for light volumes and less so for an ingestion run that pushes thousands of chunks through in one go. n8n runs on Cloud or on your own servers, which is the deciding point when the documents you are indexing cannot leave your infrastructure, since a self-hosted instance next to a self-hosted Milvus keeps everything inside. Both give you a visual canvas. Judge them on hosting, on data control and on how your volume maps to the billing model.
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