n8n workflow template · 8 nodes · Google Drive
CV Resume PDF Parsing with Multimodal Vision AI is a free n8n workflow template that starts when you run it manually in the n8n editor and works with Google Drive and Google Gemini. It has 8 nodes, 3 of them AI nodes (Structured Output Parser, Google Gemini Chat Model and Basic LLM Chain). The notes inside the workflow say: “This approach can be employed to combat "hidden prompts" planted in resumes to bypass and/or manipulate automated ATS systems using AI.” To run it, you add your own credentials for Google Drive OAuth2 and Google Gemini (PaLM) API.
The text below comes from the sticky notes inside the workflow. We removed links, code samples and anything that looks like a key or token.
This approach can be employed to combat "hidden prompts" planted in resumes to bypass and/or manipulate automated ATS systems using AI.
Join the Discord or ask in the Forum!
Read more about using Google Drive
For this demonstration, we'll pull the candidate's resume PDF from Google Drive but you can just as easily recieve this resume from email or your ATS.
It should be noted that our PDF is a special test case which has been deliberately injected with an AI bypass; the bypass is a hidden prompt which aims to override AI instructions and auto-qualify the candidate... sneaky!
Download a copy of this resume here
Read more about using Basic LLM Chain
Multimodal LLMs are LLMs which can accept binary inputs such as images, audio and/or video files. Most newer LLMs are by default multimodal and we'll use Google's Gemini here as an example. By processing each candidate's resume as an image, we avoid scenarios where text extraction fails due to layout issues or by picking up "hidden" or malicious prompts planted to subvert AI automated processing.
This vision model ensures the resume is read and understood as a human would. The hidden bypass is therefore rendered mute since the AI also cannot "see" the special prompt embedded in the document.
The text below comes from the sticky notes inside the workflow. We removed links, code samples and anything that looks like a key or token.
Read more about using Stirling PDF
AI vision models can only accept images (and sometimes videos!) as non-text inputs but not PDFs at time of writing. We'll have to convert our PDF to an image in order to use it.
Here, we'll use a tool called Stirling PDF which can provide this functionality and can be accessed via a HTTP API. Feel free to use an alternative solution if available, otherwise follow the instructions on the Stirling PDF website to set up your own instance.
Additionally, we'll reduce the resolution of our converted image to speed up the processing done by the LLM. I find that about 75% of an A4 (30x40cm) is a good balance.
For demo purposes, we're using the public online version of Stirling PDF. It is recommended to setup your own private instance of Stirling PDF before using this workflow in production.
The 8 nodes in file order (sticky notes excluded). Branches and connections determine execution order in n8n:
Configure your own credentials after importing. Create each credential in n8n, then select it in the node.
Do you need somewhere to run it? Read our guide to self-hosted n8n.
We can adapt this workflow to your apps, data and process, then deploy and support it.
Contact Us