Because sometimes, even AI needs a cheat sheet
Abstract
This project builds a Retrieval-Augmented Generation (RAG) pipeline that answers questions about Shadows of Eldoria - a fantasy story I wrote with ChatGPT. Think of it like teaching a robot to read a book and then quiz it. The pipeline uses Langflow (a visual tool for building AI workflows), Astra DB (for storage), and Ollama (to run AI models locally).
The twist? It works on your laptop. No cloud required.
Getting Started
Let's get this show on the road
Setup Instructions
- Download Ollama
Grab it from ollama.com. Then open your terminal and paste:
ollama pull llama3.2:latest ollama pull all-minilm:latest ollama serve # Keep this running!
Pro tip: If your laptop starts humming like a microwave, you're doing it right.
- Install Langflow
pip install langflow langflow run # This starts the UI at http://localhost:7860
3. Load the Pipeline
- Go to http://127.0.0.1:7860
- Import the Vector-Store-RAG.json file
- Boom. Your AI book club is ready.
How It Works
Pipeline Breakdown
- Upload Your Document
I used a 3,000-word fantasy story, but you could use anything:
- Recipes
- Legal docs
- Your grandma's shopping list
- Chop & Embed
- Cuts text into 500-character chunks (with 100-character overlaps)
- Turns words into math using Ollama's all-minilm model
Imagine Google Translate, but for human → robot.
- Store & Search
Chunks get saved in Astra DB. When you ask a question:
- It finds the 5 closest text matches
- Feeds them to Llama3.2 (the brain)
- Generates answers like a student cribbing from notes
Make It Yours
Customization options even your cat could use
Swap These Parts
- Vector Databases
Don't like Astra DB? Try:
- Chroma DB (free & local)
- Pinecone (for heavy lifting)
How? Just update the "Vector Store" node in Langflow.
- AI Models
Llama3.2 too big? Use:
- Mistral (lightweight)
- Google Gemini (if you want cloud)
Change the "OllamaModel" component with any other model component.
- Embeddings
Ollama not your jam? Switch to OpenAI embeddings. Just change the "Embeddings" node.
What's Missing?
The "uh-oh" list!
- Evaluation Framework
Right now, I judge answers by reading them. Not exactly science.
- Deployment
It lives on my laptop. To share it, I'd need to:
- Dockerize the setup
- Add user authentication
- Monitoring
What if the database crashes mid-query? No alerts exist.
Baby step: Add logging.
- Audience Clarity
Built for developers.
- Performance Stats
How fast is it? No idea.
Why This Matters
Beyond my laptop's overheating issues
Real-World Uses
-
Customer Support
Train it on your FAQ docs. -
Education
Let students query textbooks naturally. -
Personal Use
I once built a version that explains my own notes back to me.
The big picture? This is LEGO for AI pipelines.
Your Turn
What should you do next?
- Steal The Setup
Follow the instructions above.
- Break It
Try uploading a Wikipedia article instead of a story.
- Tweak
Swap Llama3.2 for Mistral. Notice speed differences?
Remember: AI isn't magic. It's just a very eager intern.
Final Thoughts
Building this felt like teaching a parrot to read. Slow. Frustrating. But when it finally answered "Who is the villain in Eldoria?" correctly? Chef's kiss.
Good AI isn't about smart models. It's about smart pipelines.
Now go build yours.
