Basic LLM Chain node calling out a model is extremly slow

[gustaff-weldon]

Bug Description

We have noticed that using Basic LLM Chain node with a model (in this case Gemini 2.0) passing down binary image for vision analysis is extremely slow. Ie. during execution a lot time is spent in llm chain node.
For comparison we have done a rest call to Gemini REST api, with the very same set of images and the results are very unfavourable for llm chain + model nodes combo.

Time to process 2 images for llm chain + gemini node:

Basic LLM Chain node: 35181ms
Model node: 4790ms

Time spent on HTTP request node uploading the very same 2 images:

HTTP Request node: 5062ms

Something is really fishy here, both use the same system prompt asking LLM to extract markdown from a given image of document.

To Reproduce

Use included workflow to process couple of invoice images and attached test invoices.

Obviously, I cannot share the documents we are processing, but sample images show 3-7x more times spent in llm chain + model nodes combo vs http request for the same images.

Extract_markdown_from_an_image.json

Expected behavior

We expected llm chain node to be equally performant to the direct REST api call to LLM API. In our tests, it is multiple times slower.

Operating System

ubuntu

n8n Version

1.80.4

Node.js Version

whatever n8n docker image comes with

Database

PostgreSQL

Execution mode

main (default)

[Joffcom]

Hey @gustaff-weldon,

We have created an internal ticket to look into this which we will be tracking as "GHC-1162"