Help Needed: Building an automation Workflow with Complex File Handling

[2musime]

Hi everyone ,

I’m working on a complex automation use case using n8n, and I’ve hit a point where I need help from more experienced builders to take it further, especially in handling files, binary data, and integrating with AI.

What I’m Trying to Build:

An AI-powered framework to receive files with different extensions(pdf,docx, XLS, PNG, JPEG,…) and I want to handle these file submissions (I am using n8n built-in form) automatically prepare professional case reports, based on uploaded documents and assignments. The workflow follows multiple structured phases:

1. *Receive files
2. Extract data tasks

Current Inputs:

Through an n8n form or webhook, I receive JSON like this:

I also push metadata about the files for the AI agent — like role (plaintiff/defendant), extraction type (OCR vs summarisation), etc.

What I Have Working:

• Form-based data collection and file intake
• JSON cleanup and metadata normalisation using Code node
• AI-ready file metadata for extraction tasks (e.g., OCR)

Where I Need Help:

• How to properly manage and reuse binary data across nodes (esp. splitting and merging)
• What’s the best pattern for processing multiple files per input row using binary data?
• Anyone implemented file-type-dependent flows (e.g., PDF → Text AI, JPEG → OCR) reliably?
• Ideas for gracefully handling large or many files in a single submission
• Any AI integration best practices from the n8n side (OpenAI vision, Claude, etc.)
• Should I offload heavy file processing (AI or storage) outside n8n?

Notes:

I’m using:

• Function & Code nodes to handle logic
• Planning to use an AI agent via GPT/OpenAI.

Any advice, example flows, or help would be appreciated.

If someone’s open to collaborating or building together — I’m open to that too !

Thanks in advance!

[RBPTG]

Have you considered using a file storage system like Amazon S3 and then using n8n to extract the contents of each file and sending only the contents through the workflow?

[Lwanda_Mhlongo]

Consider using a file storage for these files as the form might become an issue at scale.

I have built a relatively big OCR flow that handles 3000+ images a week. I’ve gotten quite good with Google Vertex.

Let me know if you’d want to discuss that too.

[2musime]

@RBPTG Thank you,
I am now going to explore it.

[2musime]

@Lwanda_Mhlongo Thank you for your suggestion. I have no experience with Google Vertex. Does it require me to shift from using n8n?

[Lwanda_Mhlongo]

No, you can do it through n8n. You feed it JPEG → It gives you the text or whatever it read in the images.