💰 How To Build A RAG Agent For Pennies (Using This n8n Workflow)

I. Introduction: RAG Just Got Ridiculously Easy (and Cheap) with an n8n Workflow

If you've built RAG agents, you know the pain. Chunking docs, generating embeddings, managing vector databases like Pinecone; it's a massive technical headache.

Google just changed the game.

With Gemini's new File Search API, you simply drop in a document and Google handles the rest: embeddings, storage, everything. You can chat with your file instantly.

The best part? It's extremely cheap.

In this guide, I'll show you how to use this API in an n8n workflow to build RAG agents without the complex data pipeline. We'll cover the build, the cost and the honest limitations you need to know.

II. What Is Gemini File Search?

Before we build, let's understand what we are actually working with.

1. How It Works

The basic process is very straightforward:

1. Upload your files to Gemini.

2. Gemini automatically chunks and embeds them for you.

3. Your n8n workflow agents can immediately use this as a knowledge base.

2. Comparison to Traditional RAG

If you have ever built a RAG agent before with a vector database, you know the pain.

• Traditional RAG Pipeline: You have to understand the file type, add metadata, add context, split documents into chunks, run them through an embeddings model, store them in a vector database, configure the search and maintain the infrastructure.

• Gemini File Search: You upload the document to Google. Google handles everything else. You plug your AI agent into the Google Store via your n8n workflow and start getting answers.

3. Why It Is Useful

• You Don't Need To: Build your own search system, set up a database, handle storage and searching manually or worry about technical infrastructure.

• All You Need: Know how to upload files through the Gemini File API and have a basic n8n workflow setup. No special technical setup is required.

III. Why Is Gemini File Search Pricing So Cheap?

Let's talk money, because this is where Gemini really shines.

1. The Cost Breakdown

• Indexing/Embedding: You are only charged when uploading. The cost is $0.15 per 1 million tokens processed.

  • Real Example: A 121-page PDF is about 95,000 tokens. That is not even one-tenth of what it would cost you 15 cents. It is extremely affordable.

• Storage: Currently, this is completely free. Regardless of how much you store, there are no storage fees at all.

• Querying: There is some limit when queries get really high but you basically only pay for the chat model usage. There is a base fee only for extremely high volume (1M+ queries/month).

2. Cost Comparison Table

Let's look at a scenario with 100 GB storage + 1 million queries per month.

The Verdict: Gemini is super cheap for a file search API. You might be able to get something cheaper if you build it all yourself from scratch but considering how easy and quick it is to set up in an n8n workflow, that cost is completely justified.

IV. What We Are Building in n8n

We are going to build a simple, four-step workflow in n8n.

1. Four HTTP Requests Overview

1. Request 1: Create Store (Folder): Think of this as creating a folder to organize your documents.

2. Request 2: Upload File: Upload the file to the Google Cloud environment (it is not in the folder yet).

3. Request 3: Move File to Store: Take the uploaded file and move it into the folder created in step 1.

4. Request 4: Query the Store: Set up the request to query the documents and get answers back.

2. The Workflow Visual

In n8n, it looks like a straight line: Create store node → Upload file node → Import file to store node → Query agent node.

After we set it up, we will run a quick evaluation to see how accurate the results actually are.

V. Step 1: Setting Up Gemini API Authentication

Honest assessment: Google's API documentation is not intuitive. It can be very confusing compared to other platforms. I will show you exactly what you need.

1. Finding the File Search API

You need to navigate to the File Search API documentation. Look for scenarios like "Directly uploading", "Importing files", "Chunking" and "Metadata".

Pro Tip: I often view the documentation as Markdown and paste it into an LLM to help me set up the requests. It is not perfect but it helps.

2. Getting Your Gemini API Key

1. Go to Google AI Studio.

2. Create your profile if you haven't already.

3. Create a new key (name it something like "test").

4. Copy the generated key.

3. Two Ways to Configure Authentication

• Basic Method (Not Recommended): Copy/paste the API key directly as a query parameter in every node. This is boring and kind of dangerous.

• Better Method (Create Authentication):

  1. In n8n, choose "Generic Credential Type" authentication type.

  2. Select the "Query Auth" option.

  3. Create a new query auth.

  4. Name field: key

  5. Value field: Paste your API key.

  6. Name the credential: (e.g., "Gemini File Search").

  7. Save. Now you can just select this credential for every node in your n8n workflow.

VI. Step 2: Creating a File Search Store

First, we need a place to put our files.

1. Configure the HTTP Request

• Endpoint: Find the "Create Store" POST request in the documentation and paste the URL into your n8n node.

• Authentication: Select the "Gemini File Search" credential we just made.

• Headers: Set Content-Type to application/json.

2. The Body Configuration

This is the most important part. In the JSON body, you need to choose a name for your file store.

• Example: {"displayName": "Test"}

3. Execute and Pin

Click "Execute". You will get a successful response from Google containing the name of the file store (it will look like an ID) and the display name.

Critical: Pin this data. You will need the store name ID for the next steps.

VII. Step 3: Uploading the File to Gemini

This is a two-step process. First, we upload it to the cloud. Then, we move it into the store.

Before uploading the file, we need a file, right?

• Let’s get started by adding a Form Trigger.

• Then customize it with a File field, which allows us to upload multiple files.

1. Setting Up the Upload Request

• Endpoint: Look for "Initiate resumable upload to store" in the documentation. It is a POST request ending in /upload/v1beta/files.

• Authentication: Use your saved credentials.

2. Sending the File

• Body Configuration: Choose the "n8n Binary File" option.

• Input Data: Look at your trigger data (where your file came from). Find the binary file name (e.g., "file") and type that name into the "File Name" field in the HTTP request.

3. Execute the Upload

Click "Execute". You will get a success message with details like the file name, MIME type, size and URI.

Important: This file is now in the Google environment but it is just floating there. It has an expiration time. If you don't move it to a folder (store), it will be deleted.

VIII. Step 4: Moving the File Into the Store

Now we connect the file to the store we created in Step 2.

1. Configure the Import Request

• Endpoint: Find "Import files into file store" in the documentation.

  • Critical Detail: The URL includes a variable for the store name: .../{storeName}:importFile.

• How to Configure: Go back to your "Create Store" node (Step 2), copy the store name ID from the pinned data and paste it into the URL where {storeName} appears.

2. JSON Body

You need to tell it which file to import.

• In the JSON body, you need to send the resource name of the uploaded file.

• Drag the name (or uri) field from the "Upload File" node's output (Step 3) into the JSON body value.

3. Execute the Import

Click "Execute". You will get a success message. The file has been moved into the folder and is now ready to be queried by your n8n workflow.

IX. Step 5: Setting Up the RAG Agent

Now for the fun part: chatting with your data.

1. Configuring the Knowledge-Based Tool

We will use another HTTP Request tool to query the store.

• Endpoint: Use the "Generate content" request for the Gemini 2.5 Flash model.

• Method: POST.

2. Critical Configuration: Store Name

In the JSON body of this request, you need to tell the agent which store to search.

• Paste the store name ID (from Step 2) into the appropriate field in the JSON body. Now the agent knows where to look.

3. Setting Up the Query

We use the fromAI function in n8n.

• Query: Set the query value to "the question that the user needs an answer to".

• The AI agent controlling this tool will automatically fill in this query based on the user's chat input.

4. Prompt Engineering

We need to give the agent clear instructions.

You are a helpful RAG agent. Your job is to answer the user's question using your Knowledge Base tool to make sure all of your answers are grounded in truth. Please cite your sources when you're giving your answers. 

When you are sending a query to the knowledge base tool, only send over text. No punctuation, quotation marks or new lines. 

Why This Matters: The instruction about punctuation keeps the JSON structure intact, preventing errors.

X. Testing the Agent: Does It Work?

I uploaded the Rules of Golf PDF as a test document.

Test Question: "What happens if your club breaks during the round?"

The Result:

• Short Answer: "If a club breaks during the round because of normal play, you may continue to use the damaged club, have it repaired or replace it... But if it breaks for non-normal reasons, you may not use that club…"

• Source Cited: "Rules on clubs and damage".

Verdict: It successfully handled the document and gave a correct, cited answer.

Testing with Multiple Documents

I added a second file: an Nvidia press release (completely different topic).

Test Question: "Give me a Q1 2025 fiscal summary for Nvidia".

The Result:

• Summary: It provided total revenue ($26 billion), data center revenue ($22 billion) and other key stats.

• Source Cited: "NVIDIA Q1 FY2025 financial results press release (quarter ended April 28, 2024) retrieved from Knowledge Base".

Verification: I cross-referenced it with the actual document and the information was correct.

XI. Evaluating Accuracy

I ran a more rigorous evaluation. I had three documents in the store:

1. Rules of Golf (22 pages)

2. Nvidia Announcement (9 pages)

3. Apple 10-K (121 pages)

Total: Nearly 200 pages across three completely unrelated topics.

I asked 10 difficult questions spanning all three PDFs.

• Overall Score: 4.5 out of 5 for correctness.

• Individual Scores: Most were 5/5.

Assessment: For a system with no special processing and minimal prompting, these results are impressive.

XII. What Limitations Should You Know Before Using File Search?

While it is great that you can drop in files easily, there are important limitations you must understand when building this n8n workflow.

Consideration #1: Duplicate Data Management

• The Problem: Google isn't tracking versions. If you upload an updated version of a file, it creates a duplicate.

• The Impact: Your database explodes with duplicate data, which lowers the agent's response quality.

• The Solution: You must build your own file management system. Track versions manually and delete old files before uploading new ones.

Consideration #2: Garbage In, Garbage Out

• The Reality: Gemini has some OCR capability but it can't fix a super messy document. If your source material is bad, the answers will be bad.

• The Solution: You might need to pre-process or clean up messy scanned documents before dropping them into Gemini.

Consideration #3: Chunk-Based Retrieval Limitations

• When Semantic Search Works: It is great for finding a "needle in a haystack", specific facts like "What is Rule 28?"

• When It Fails: It struggles with holistic understanding.

  • Example: I asked, "How many total rules are in your PDF?" It answered "Five". The reality is there are many more but the chunk-based retrieval couldn't "see" the entire document at once to count them.

• The Bottom Line: Unless you use advanced metadata tagging, this is not the best path for comprehensive document analysis or summarization.

Consideration #4: Security and Privacy

• The Reality: Your documents are uploaded and stored on Google's servers.

• What This Means: Do not upload sensitive PII (Personally Identifiable Information) or highly confidential trade secrets unless you are comfortable with Google processing and indexing that data.

• Compliance: Always consider GDPR, HIPAA and CCPA regulations.

XIII. Conclusion: Is Gemini File Search Worth It?

For most use cases, the answer is Yes.

Let's look at the math:

• Setup Time: 30 minutes.

• Cost: Essentially free for moderate usage.

• Quality: 4.5/5 accuracy with minimal tuning.

• Simplicity: Just four HTTP requests.

Compare this to traditional RAG, which takes hours or days to set up, costs significantly more and has much higher complexity.

Who Should Use This:

• Developers testing RAG concepts.

• Small businesses with basic Q&A needs.

• Content creators organizing knowledge.

• Anyone starting with RAG agents who is budget-conscious.

Who Should Avoid This:

• Enterprises with strict compliance needs.

• Scenarios requiring full document context (like "summarize this whole book").

• Highly sensitive data.

Gemini's file search API dramatically lowers the barrier to entry for RAG agents. It is not perfect but the combination of ease, speed and cost makes it a compelling choice for 90% of use cases.

Stop overthinking your RAG infrastructure and start building n8n workflows that actually deliver value.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• Bulletproof Your n8n AI Agents: 7 Pro Settings Most People Miss!*

• Stop Using n8n Templates... Until You Read This!*

• Google Antigravity: The AI That "Watches" You Code

• 20 New AI Tools That Are FREE... For Now

• 12 Mind-Blowing Apps Gemini 3 Built From Scratch

  *indicates a premium content, if any