Building a Deep Research Workflow with n8n: My Setup, Tools, and Learnings

Lately, I’ve been experimenting with automating deep research using n8n. The goal was simple: type a prompt, wait a few minutes, get back a structured research report with summaries and sources — no fluff, no extra work.

It took time to get right. I watched a bunch of research automation videos, borrowed what made sense, and hacked the rest together. Now it works reliably.

Here’s a breakdown of how I did it, what tools I used, and what came out of it.

What I Wanted

A prompt like:
"Clinical Communication Market in the US"

In return, a structured report within 8 minutes. Ideally with bullet-point insights, links to sources, and export options.

Tools I Used

OpenRouter

This is where the prompts go. I used OpenRouter to access GPT models that could break down broad topics into sub-queries and generate clean summaries from source material.

Tavily

A powerful search API that goes beyond just returning links. It gives back structured results — perfect for feeding into the summarization step. Each subquery created by the AI is sent here.

Google Sheets

I used Google Sheets as a lightweight, shareable database. The workflow writes summaries, links, and tags into the sheet so I can review everything easily or hand it off to someone else.

Template.io

To create a polished final report, I integrated Template.io. It takes the structured content from Google Sheets and turns it into a clean, well-formatted PDF. I set up a basic report template and let the workflow populate it dynamically.

How It Works

  1. Input prompt
    I enter a broad research topic in n8n

  2. Prompt refinement via OpenRouter
    The AI breaks it into focused subtopics or search queries

  3. Query execution via Tavily
    Each query is sent to Tavily for fresh, relevant results from the web

  4. AI summarization
    Results from Tavily are summarized using GPT via OpenRouter again

  5. Store in Google Sheets
    Each summary, source link, and relevance tag is saved row by row

  6. Generate PDF with Template.io
    Optionally, a final report is generated using the data in the sheet and a preset PDF template




What It Produced

I tested the workflow with the prompt:
"Clinical Communication Market in the US"

It took around 8 minutes to run and cost me roughly $0.50 USD total in API calls.

The result was a clean, structured report with real-time insights from multiple sources. Not perfect, but good enough for internal research, presentations, or early market validation.

Try It Out

Here’s the output if you want to see what it generated → the Pdf file. 

 

If you're curious about the workflow, happy to share it. And if you've built something similar or have ideas to make this better, I’d love to trade notes.


Comments

Post a Comment

Popular posts from this blog

Running Llama LLM locally without a GPU

Image
A few mornings (months!) ago, I read that Meta has made LLaMA available to the public and you can host the LLMs on your local machine. I have always associated AIs with GPUs, and GPUs are expensive. So the extent of my understanding of AI was that it was expensive—and ChatGPT is free. Then, out of sheer curiosity, I decided to search if I would be able to run a local LLM on my i5, 16 GB Windows machine. Voila, apparently I could do that in less than half an hour. Trust me, I was neck-deep in work when I decided to take this route. So there you go—a step-by-step (no code required!) way to run a LLaMA 3.2:1b model on your basic Windows machine. I’m not sure if anything above this model will work with such a configuration, but LLaMA 3.2:1b does. Step 1:  Installing Docker Desktop  If you don’t have Docker Desktop, install it. It’s just a few “Next” clicks and one “Install” button. You can download it from here —just choose your desktop configuration, download, and install. Now,...
Read more