Home Projects Blog What I Do Experience Contact
Book a call
Back to Blog

RAG in 2025: Traditional vs Modern,
Workflow vs Agentic

You've probably heard the term RAG thrown around a lot lately. This guide breaks it down from the ground up — no assumed knowledge needed. Whether you're a business owner evaluating tools, a developer choosing a stack, or just curious how AI assistants actually work with your documents, this will give you a clear picture.

1. What is RAG?

AI language models like ChatGPT are incredibly smart — but they only know what they were trained on. They don't know about your company's documents, your internal policies, or anything that happened after their training ended. And when they don't know something, they sometimes make up a confident-sounding answer. This is called hallucination.

RAG (Retrieval-Augmented Generation) solves this. Instead of letting the AI guess, RAG first searches your actual documents, finds the most relevant information, and hands it to the AI before it answers. The AI then responds based on what it just read — not what it vaguely remembers.

Think of it like this

Imagine asking someone a question during an exam. Without RAG, it's a closed-book exam — they can only answer from memory. With RAG, it's an open-book exam — they look up the answer in the right document before responding. Much more reliable.

Every RAG system works in three simple steps:

  1. Store — your documents are broken into small sections and saved in a searchable database.
  2. Search — when someone asks a question, the system finds the most relevant sections from your documents.
  3. Answer — the AI reads those sections and uses them to write a grounded, accurate response.
PHASE 1 — INDEXING (happens once) PHASE 2 — QUERY TIME (every question) 📄 Your Documents split 🔢 Embed convert to vectors store 🗄 Vector Database all doc chunks, searchable by meaning 💬 User asks a question search for similar 📋 Top chunks most relevant passages retrieve 🧠 LLM reads chunks + answers also passes question ✅ Answer
Fig 1 — RAG in two phases: index documents once, then search + generate on every question

2. Traditional RAG vs Modern RAG

Traditional RAG — Simple and Effective

The original approach to RAG is straightforward. You upload documents, they get split into chunks, the system searches for the most relevant chunks when someone asks a question, and the AI answers using those chunks. One question in, one answer out.

This works well for most everyday use cases — internal FAQ bots, searching through a company handbook, or answering questions about a product. If your questions are simple and your documents are well-organised, traditional RAG is all you need.

When traditional RAG is enough

Building a chatbot to answer staff HR questions? Searching a few hundred product documents? Traditional RAG is the right choice. Don't overcomplicate it.

Modern RAG — Smarter Searching

Modern RAG adds intelligence to the search process. Instead of doing one search and hoping for the best, it can search multiple times, re-check its own answers, and combine information from different places. Here's what that looks like in practice:

Which one do you need?

Traditional RAG simple, single-pass 💬 Question 🔍 Search once 🧠 LLM answers ✅ Answer ❌ can miss info if search fails Modern RAG multi-step, self-checking 💬 Question 🔍 Hybrid search (semantic + keyword) 📊 Rerank results 🧠 LLM answers self-check + retry
Fig 2 — Traditional RAG: one search pass. Modern RAG: hybrid search, re-ranking, and self-verification loops

3. Two Ways to Build a RAG System

Once you know you need RAG, the next question is: how do you build it? There are two main approaches.

Workflow RAG is like a recipe. You define every step in advance — load the document, search the database, ask the AI, return the answer. It always follows the same steps in the same order. Tools like N8N let you build this visually, without writing code.

Agentic RAG is more like hiring a smart researcher. Instead of following fixed steps, it reads the question, decides what to look up and where, goes back and forth between sources if needed, and puts together a complete answer. It reasons through the problem rather than following a script.

The key difference

Workflow RAG always does the same thing, step by step. Agentic RAG thinks about the question first, then decides how to answer it. For simple questions, the workflow is faster. For complex ones, the agent wins every time.

Workflow RAG (N8N) follows a fixed recipe, every time Step 1: Receive question Step 2: Search database Step 3: Ask LLM Step 4: Return answer 🔒 Same 4 steps for every question Can't adapt to complex queries Agentic RAG reasons first, then decides what to do 🧠 Agent reasons about the question 💬 Any question 🔍 Search docs 📊 Query data table 💡 Reason can loop back if answer needs more info ✅ Synthesises final answer
Fig 3 — Workflow RAG locks you into fixed steps; Agentic RAG reasons about the question and chooses its tools dynamically

4. Workflow RAG with N8N

N8N is a visual workflow tool — you connect blocks together on a canvas to build automated processes, no coding required. It has built-in AI blocks, which makes it a popular choice for building RAG pipelines quickly.

A typical N8N RAG pipeline looks like this:

How N8N Workflow RAG Works ① INGESTION — document upload 📄 Document comes in ✂️ Load & split it 🔢 Convert to vectors 💾 Save to database ② QUERY TIME — every question, in order 💬 User asks question 🔍 Search database 🤖 AI reads & answers Response sent Every question follows the exact same steps — no reasoning, no flexibility
Fig — N8N Workflow RAG: a fixed two-phase pipeline (ingestion + query). Every question walks the same path, in the same order.

The biggest advantage of N8N is that non-technical team members can understand and even modify the workflow. It also connects easily to tools your team already uses — Slack, email, CRMs, Google Sheets — so you can wire AI answers directly into existing processes.

The main limitation is that it's rigid. Every step is fixed in advance. If a question needs a different kind of search or requires combining answers from two different sources, N8N can't figure that out on its own — you'd have to manually build every possible variation.

Best for: customer support bots, internal knowledge bases, document Q&A for small teams, and any workflow where a non-technical person needs to manage the setup.

5. Agentic RAG on AWS — A Real System

This is an architecture from a production system I built on AWS. It handles much more complex questions than N8N can — because instead of following fixed steps, it has a team of specialised AI agents that work together.

Step 1 — Uploading documents

DOCUMENT INGESTION — what happens automatically when you upload a PDF 📄 PDF Uploaded to AWS S3 ① upload triggers Lambda Function splits into chunks ② auto-process embed 🧠 Bedrock Embeddings converts text → vectors ③ vectorise store 🗄 Vector Database RDS Aurora (pgvector) ④ ready to search 🔒 Everything stays inside a private AWS VPC — no document leaves your account
Fig 5 — Upload a PDF once; the pipeline automatically chunks, embeds, and stores it — no manual steps

When you upload a PDF, the system automatically processes it in the background. It splits the document into sections, converts each section into a format that's optimised for searching, and stores everything in a secure private database. No manual steps needed.

Step 2 — Answering questions

HOW A QUESTION GETS ANSWERED 👤 User sends a question 🌐 API Gateway public entry point public internet enters VPC 🔒 Private AWS VPC all actions logged → CloudTrail 🎯 Supervisor Agent reads question · decides which agents to call 📄 Document Search finds relevant PDF sections ↳ searches Vector Store (S3) 📊 Data Agent counts, figures, lookups ↳ queries RDS Aurora ✍️ Synthesis Agent combines all results ↳ writes the final reply ✅ Grounded answer returned
Fig 6 — API Gateway is the only public entry point; the Supervisor, all agents, and datastores run entirely inside the private VPC

When a user asks a question, a Supervisor Agent reads it and decides the best way to answer. It can call on three specialist agents depending on what's needed:

For a simple question like "What is our refund policy?", only the Document Search Agent gets called. For a complex question like "How many support tickets were raised last month and what were the most common issues?", all three agents work together — the Data Agent counts the tickets, the Document Search Agent finds patterns in the text, and the Synthesis Agent writes up the combined answer.

All of this runs privately inside AWS — nothing is exposed to the public internet, and every action is logged automatically for security and compliance.

6. Other Agentic RAG Options

AWS isn't the only way to build agentic RAG. Here's a quick look at the main alternatives:

Platform How much control How fast to set up Cost Best for
AWS Bedrock Agents High Takes time Pay per use Enterprise teams on AWS
LangGraph + Lambda Maximum Medium effort Low Developers who want full control
OpenAI Assistants Limited Very fast Moderate Quick prototypes
Vertex AI Agents High Medium effort Pay per use Teams already on Google Cloud

7. N8N vs Agentic RAG — Side by Side

What matters N8N Workflow RAG Agentic RAG (AWS / LangGraph)
Can it reason through complex questions? No — fixed steps only Yes — thinks before acting
Multiple AI agents working together? Manual setup only Built-in
How hard is it to set up? Easy Requires technical team
Handles large document libraries? Small to medium Yes, enterprise scale
Cost Low / free if self-hosted Pay per use
Can you customise everything? Limited to what nodes support Full control
Does it check its own answers? Not really Yes — self-correction built in
Maintenance Low — visual interface Needs ongoing DevOps
Best for Automations & integrations Complex Q&A over large data

8. When N8N is the Right Choice

N8N works best when:

9. When Agentic RAG is the Right Choice

Agentic RAG is the right choice when:

10. Security — Which is Safer?

Security is one of the biggest differences between these two approaches. For businesses handling sensitive data, it matters a lot.

Security area N8N RAG AWS Agentic RAG
Is your data isolated from the internet? Not by default Yes — runs in a private network
Who can access what? API keys only Granular permissions per service
Where are passwords and credentials stored? N8N's encrypted store AWS Secrets Manager — never in code
Is stored data encrypted? Depends on your setup Always encrypted at rest
Is data encrypted while moving? Basic HTTPS Encrypted + never leaves AWS network
Is there a full activity log? Not built-in Yes — every action is recorded
Automatic threat detection? No Yes — AI-based monitoring
Suitable for regulated industries? Internal tools only Yes — GDPR, HIPAA compliant
Overall rating ⭐⭐
Fine for internal use
⭐⭐⭐⭐⭐
Enterprise ready
Bottom line on security

AWS is significantly more secure. It keeps your data private, logs every action, and encrypts everything automatically. N8N is perfectly fine for internal tools where sensitive customer or business data isn't involved. But if you're handling confidential documents, customer records, or regulated data — AWS is the only responsible choice.

11. Which Approach is Right for You?

The goal isn't to build the most sophisticated AI system — it's to build the right one for your situation. A simple tool that works reliably is always better than a complex one that's hard to maintain or breaks under pressure. Know what you need, match it to the right approach, and build from there.

Start simple. Scale up only when the problem actually demands it.