PGVector Store Node: Error inserting expected 1024 dimensions, not 256 when using self-hosted embedding model

[BillRaymond]

Bug Description

I’m running a RAG workflow in self-hosted n8n and consistently receive this error from the Postgres PGVector Store node:

Error inserting: expected 1024 dimensions, not 256

My setup:

• n8n self-hosted (v1.118.1)
• Postgres 18 with pgvector extension
• chunks table embedding column: vector(1024)
• Embedding model: text-embedding-bge-m3 (running via LM Studio using OpenAI compatibility)
• n8n node: Embeddings OpenAI (connected to LM Studio)
• Dimensions explicitly set to 1024

Confirmed that:

• The embedding output array has exactly 1024 floats (verified with length in a Function node)
• Postgres column is vector(1024)
• Yet n8n throws expected 1024, not 256 during insert.

Also confirmed that a basic HTTP POST will return a 1024 vector:

curl -X POST "http://my-lm-studio-server-using-openai-compatibility/v1/embeddings" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-bge-m3",
    "input": "My text to embed"
  }'

To Reproduce

1. Create a Postgres (with PGVector) table that contains a field with an embedding vector (1024).
2. Create a workflow that starts with a Set fields node and contains some text you want to embed. Output it as data
3. Connect a Postgres PGVector Store node to the Set fields node

• Credential to connect with: your-postgres-with-vector-database-connection
• Operation mode: Insert Documents
• Table name: the table you created in Step 1
• Embedding batch size: 200 (default)

3. Connect the Postgres PGVector Store node to Embeddings OpenAI (Embeddings)

• Credential to connect with: An LM Studio server, which by default is compatible with OpenAI APIs
• Model: text-embedding-bge-m3
• Dimensions: 1024

4. Connect the Postgres PGVector Store node to Default Data Loader (Document)

• Type of data: JSON
• Mode: Load Specific Data
• Data: Text data from a previous node you want to embed
• Text splitting: Custom

5. Connect the Default Data Loader to the Token Splitter node

• Chunk size: 800
• Chunk overlap: 100

6. Execute the workflow

Error: Problem in node ‘Postgres PGVector Store‘
Error inserting: expected 1024 dimensions, not 256

Refer to the Bug Description to note that a standard CURL POST test yields a 1024-dimensional response from the same endpoint, which is why I believe this to be an n8n bug.

Expected behavior

The text embedding returns a 1024-dimensional response, just as with a standard POST event to the same URL.

Debug Info

Debug info

core

• n8nVersion: 1.118.1
• platform: docker (self-hosted)
• nodeJsVersion: 22.21.0
• nodeEnv: production
• database: sqlite
• executionMode: regular
• concurrency: -1
• license: enterprise (production)
• consumerId: be538b36-658e-435e-8683-d734f810c123

storage

• success: all
• error: all
• progress: false
• manual: true
• binaryMode: memory

pruning

• enabled: true
• maxAge: 336 hours
• maxCount: 10000 executions

client

• userAgent: mozilla/5.0 (macintosh; intel mac os x 10_15_7) applewebkit/537.36 (khtml, like gecko) chrome/142.0.0.0 safari/537.36
• isTouchDevice: false

Generated at: 2025-11-06T00:33:00.432Z

Operating System

Debian GNU/Linux 13 (trixie)

n8n Version

1.118.1

Node.js Version

2.21.0

Database

SQLite (default)

Execution mode

main (default)

Hosting

self hosted

[n8n-assistant]

Hey @BillRaymond,

Thank you for reaching out! We’ve received your issue and are looking into it. To help us track this internally, we’ve created a Linear ticket with the reference: "GHC-5378".

We’ll keep you updated as we make progress, but if you have any additional details or context to share, feel free to add them here - it’s always helpful!

Thanks for your patience and for bringing this to our attention.

---

For INTERNAL use only: GHC-5378

[BillRaymond]

For some additional color, I have learned that the embedding issue is not just related to write methods, but also to read methods. I've attached a copy of my workflow.

Below is a copy of the error message. Please note, I anonymized the data as follows:

1. Environment-specific paths
2. Removed file hashes
3. Identifiable information

{
  "errorMessage": "different vector dimensions 256 and 1024",
  "errorDetails": {},
  "n8nDetails": {
    "n8nVersion": "1.118.2 (Self Hosted)",
    "binaryDataMode": "default",
    "stackTrace": [
      "error: different vector dimensions 256 and 1024",
      "    at /path/to/n8n/node_modules/pg-pool/index.js:45:11",
      "    at processTicksAndRejections (node:internal/process/task_queues:105:5)",
      "    at ExtendedPGVectorStore.searchPostgres (/path/to/n8n/node_modules/@langchain/community/dist/vectorstores/pgvector.cjs:635:28)",
      "    at ExtendedPGVectorStore.similaritySearchVectorWithScore (/path/to/n8n/node_modules/@n8n/n8n-nodes-langchain/nodes/vector_store/VectorStorePGVector/VectorStorePGVector.node.ts:212:10)",
      "    at handleRetrieveAsToolExecuteOperation (/path/to/n8n/node_modules/@n8n/n8n-nodes-langchain/nodes/vector_store/shared/createVectorStoreNode/operations/retrieveAsToolExecuteOperation.ts:62:14)",
      "    at ExecuteContext.execute (/path/to/n8n/node_modules/@n8n/n8n-nodes-langchain/nodes/vector_store/shared/createVectorStoreNode/createVectorStoreNode.ts:300:19)",
      "    at WorkflowExecute.executeNode (/path/to/n8n/node_modules/n8n-core/src/execution-engine/workflow-execute.ts:1093:8)",
      "    at WorkflowExecute.runNode (/path/to/n8n/node_modules/n8n-core/src/execution-engine/workflow-execute.ts:1274:11)",
      "    at /path/to/n8n/node_modules/n8n-core/src/execution-engine/workflow-execute.ts:1708:27",
      "    at /path/to/n8n/node_modules/n8n-core/src/execution-engine/workflow-execute.ts:2324:11"
    ]
  }
}

Here, you can see the same error:

❌ Problem in node ‘Postgres PGVector Store‘
different vector dimensions 256 and 1024

Here are the embeddings, showing that I have set the node to 1024, but n8n defaults to 256:

Here are screenshots of the embedding field's definitions:

[jabbson]

Have you tried using native OpenAI Embedding model? I wonder if this is where the problem is coming from.

[jabbson]

Error inserting: expected 1024 dimensions, not 256
while inserting the document sounds like the expectation (the database) is 1024, but provided vector was 256.

I just tried with regular openai emb model and it went as smooth as one could've hoped for:

[BillRaymond]

I have tested quite a bit and am not sure exactly where the breakdown is, but I can guarantee both match.

[BillRaymond]

Have you tried using native OpenAI Embedding model? I wonder if this is where the problem is coming from.

No, I am using an open-weight embedding model. I was using the OpenAI-compatible API. I can confirm this is an issue with the n8n node. I am pretty sure what is happening here is that while it knows the setting is 1024, it does not recognize that as a value to use when running against the model.

[hubin858130]

I am using the bge-m3 model, have the same issue.

[hubin858130]

I have identified the issue, it is not a bug,We need to set some parameters in n8n.

[tamy-commits]

Hi @BillRaymond

Thanks for reporting this issue! We’ve attempted to reproduce the problem, but unfortunately, we weren’t able to replicate it. It’s possible there was some environmental or other variable that prevented us from seeing the same behaviour.

If you’re still encountering this issue, please provide any additional details or steps to reproduce it, and we’d be happy to investigate further. Thanks for your understanding!

#closed:cant-reproduce