n8n to Weaviate: How to Sync Your Data

Syncing n8n to Weaviate turns your workflow execution history into a searchable vector index, which is the groundwork for asking questions about your automation layer in natural language rather than SQL.

Check your Weaviate version before anything else. Weaviate's own documentation states that the Airbyte integration uses the v3 Python client and is only compatible with Weaviate Database versions below 1.24. Airbyte's docs give a minimum of 1.21.2. That leaves a supported window of roughly 1.21.2 through 1.23.x. If your cluster is on 1.24 or later, which most current deployments are, this connector is not going to work for you and no amount of configuration will change that.

The rest of this guide assumes you have confirmed your cluster falls inside that window, or that you are running one deliberately for this purpose.

n8n to Weaviate at a glance:

Why put n8n executions in a vector database?

Execution records carry error messages, payloads, and node output, which is exactly the kind of semi-structured text that similarity search handles better than exact matching. Indexing them lets you ask what else failed like this, or surface past runs resembling one you are debugging, without knowing in advance which error string to grep for.

It is also the ingestion half of a retrieval-augmented generation setup over your own operational history. If you want counts, failure rates, or durations rather than semantic search, a warehouse is the better destination and the vector index is unnecessary overhead.

What does the n8n source give you?

One stream: executions. The n8n source reads from the n8n API and exposes the record of workflow runs. Workflow definitions, credentials, users, and tags are not available.

It also supports full refresh only, with no incremental mode, so every sync re-reads the entire execution history. In a vector pipeline this matters more than usual, because re-reading means re-embedding, and embedding costs money per call. A busy n8n instance synced hourly will generate a substantial OpenAI or Cohere bill for data that has not changed. Daily is a more defensible cadence, and overwrite is the sync mode to use.

The connector also carries a low sync success rate in Airbyte's own metadata, so monitor it rather than assuming silence means success.

How does the Weaviate destination work?

The Weaviate destination runs a three-stage pipeline rather than a simple load. Processing splits records into chunks that fit a context window, and decides which fields are text and which are supplementary metadata. Embedding converts that text into vectors. Indexing stores them for similarity search.

That middle stage is the configuration decision that matters most.

• OpenAI produces embeddings with text-embedding-ada-002 and needs an API key with billing enabled.
• Cohere uses embed-english-light-v2.0 at 1024 dimensions.
• From field loads pre-calculated embeddings already present in your records. The field must be a JSON array of numbers. This does not apply to raw n8n executions, which carry no vectors.
• No embedding is for clusters where you have configured a vectorizer on the class itself. Be aware that users have reported this option failing at the connection test with an abstract method error, so test it before committing to that design.
• Fake generates random vectors and exists to test a pipeline without incurring embedding costs. Use it for your first run.

How do you set it up?

In n8n, open Settings, then API, and create an API key. Configure the source with that key and your instance host. On the Weaviate side you need the cluster URL, credentials as either an API token or username and password, and an API key for whichever embedding service you chose. Airbyte Cloud connects to Weaviate over TLS only. If your organisation filters by IP, add the Airbyte Cloud IP addresses to the relevant allow lists.

In the Processing section, nominate which execution fields carry the text worth embedding and which are metadata for filtering. Being selective here directly reduces your embedding bill, since every field you mark as text becomes tokens you pay to vectorise on every full refresh.

Each stream is indexed into its own class derived from the stream name, and Airbyte creates the class if it does not already exist using your configured vectorizer settings. Objects get a generated UUID, with the originating Airbyte record identifier stored in _ab_record_id. Use that property when you need to trace a Weaviate object back to a specific execution.

Frequently asked questions

Will this work with my Weaviate 1.24 or newer cluster?

No. Weaviate's documentation states the Airbyte integration uses the v3 Python client and is compatible only with database versions below 1.24. Confirm your version before investing time in configuration, and check both vendors' current documentation in case the client has since been updated.

How do I avoid a large embedding bill?

Sync daily rather than hourly, be selective about which fields you mark as text, and use the Fake embedding option while you are still testing the pipeline shape. Because the source is full refresh only, every sync re-embeds everything.

Can I bring my own embeddings?

Yes, through the From field option, provided the vectors already exist in your records as a JSON array of numbers. Raw n8n execution records do not contain vectors, so this only applies if you are transforming them upstream first.

Which sync mode should I use?

Overwrite. The n8n source cannot do incremental, so append would accumulate a fresh copy of every execution, along with a fresh embedding cost, on every run.

Get your n8n data into Weaviate

Confirm the version ceiling first, test with fake embeddings, then switch to a real provider once the shape is right. Want counts and durations rather than semantic search? See our guides to n8n to ClickHouse and n8n to Convex.

Start syncing now →