Build a $10,000 RAG system using Gemini + Claude Code

STEP 1. Understand what you’re building

You are creating a system where:

→ Text, images, videos, documents live in one database

→ All data gets embedded into the same vector space

→ AI retrieves the most relevant pieces before answering

This is RAG

Retrieval Augmented Generation

✦ Key shift

Old systems handled text only

New systems handle meaning across formats

STEP 2. Set up your tools

You need 3 things:

1. Gemini API → Used for embeddings → Get from Google AI Studio
2. Pinecone → Your vector database → Stores embeddings
3. OpenRouter or model provider → Used for chat responses
4. Visual Studio Code → Your working environment
5. Claude Code → Builds everything for you

STEP 3. Create your project

→ Open VS Code

→ Install Claude Code extension

→ Open a new folder

Now open Claude Code panel

Switch to plan mode

Paste documentation link for Gemini embeddings

Then prompt:

“Build a multimodal RAG system using Gemini Embedding 2 and Pinecone.

Create env file placeholders for API keys.

Support text, images, and videos.”

Claude Code will generate:

→ Project structure

→ Dependencies

→ Step-by-step plan

Accept it

STEP 4. Add API keys

In your env file, add:

→ Gemini API key

→ Pinecone API key

→ OpenRouter or model key

Save the file

That’s it for setup

STEP 5. Add your data

Create a “data” folder

Drop in anything:

→ PDFs

→ Images

→ Videos

→ Text files

No need to organize perfectly

The system handles classification

STEP 6. Run ingestion

Prompt Claude Code:

“Process all files and store embeddings in Pinecone.

Then build a simple chat app.”

What happens behind the scenes:

→ Files get chunked

→ Gemini creates embeddings

→ Data stored in Pinecone

→ Metadata added

✦ This is where older tools like n8n get messy

Manual chunking

Separate pipelines

Frequent failures

Here, it runs in one flow

STEP 7. Test your system

Claude Code builds a local app

You open localhost

Now test queries:

→ “How do I clean the filter?”

↳ Returns steps + images from PDF

→ “What are the parts?”

↳ Pulls multiple sections + diagrams

→ Upload an image

↳ Finds similar entries in database

STEP 8. Improve retrieval quality

By default:

→ Images and videos are stored as descriptions

To improve:

Ask Claude Code:

“Add better metadata descriptions for images and videos

Update app to display media inline”

Now your system:

→ Shows images

→ Plays videos

→ Gives richer results

STEP 9. Understand limitations

Current constraints:

→ Video length limit around 120 seconds

→ Image batch limits

→ Quality depends on metadata

✦ Important

Better descriptions = better retrieval

STEP 10. Real use cases

1. Instruction manuals → Chat with complex PDFs → Get visual answers
2. Service businesses → Upload project images → Retrieve similar jobs with pricing
3. Internal knowledge bases → Mix documents, videos, images → One unified search

STEP 11. What changed

Before:

→ Complex n8n pipelines

→ Manual configuration

→ Fragile systems

Now:

→ Describe system in plain language

→ AI builds it

→ You refine outputs

Mini insight

This build took under 30 minutes

Earlier versions took hours or days