Anime Recommender System using LLMs & Kubernetes

A Large Language Model (LLM)-powered Anime Recommender System built using LangChain, GROQ API, ChromaDB, and deployed via Kubernetes on a GCP VM with Docker & Minikube.

---

Features

• Personalized anime recommendations powered by GROQ LLM
• Retrieval-Augmented Generation (RAG) using Chroma vector DB
• Prompt-driven contextual search over anime metadata
• Built-in Streamlit app UI
• End-to-end containerized deployment using Kubernetes on GCP
• Monitoring with Grafana + Helm

---

Project Structure

.
├── app/
│   └── app.py                     # Streamlit frontend app
├── config/
│   └── config.py                  # Loads .env vars (e.g., GROQ API Key)
├── pipeline/
│   ├── build_pipeline.py          # Builds vector DB from anime dataset
│   └── pipeline.py                # Runs LLM inference pipeline
├── src/
│   ├── data_loader.py             # Reads and processes raw anime CSV data
│   ├── vector_store.py            # Embeds + stores vectors into ChromaDB
│   ├── prompt_template.py         # Custom prompt template for LLM
│   └── recommender.py             # GROQ LLM + LangChain RetrievalQA logic
├── utils/
│   ├── custom_exception.py        # Centralized error tracking
│   └── logger.py                  # File-based logging support
├── Dockerfile                     # Docker build config
├── llmops-k8s.yaml                       # Kubernetes Deployment & Service spec
├── .env                           # Environment secrets (GROQ API Key, etc.)
└── README.md                      # You're here!

---

How It Works

Step-by-step Pipeline

1. Data Processing (data_loader.py):

   • Reads raw anime CSV
   • Combines metadata into a text corpus
   • Saves cleaned data

2. Embedding & Storage (vector_store.py, build_pipeline.py):

   • Splits text into chunks
   • Converts chunks into embeddings using HuggingFace
   • Saves them in Chroma vector DB

3. Prompt Engineering (prompt_template.py):

   • Creates a custom prompt template that instructs the LLM to recommend anime titles

4. LLM-powered Retrieval (recommender.py, pipeline.py):

   • Uses LangChain RetrievalQA with GROQ LLM
   • Fetches top relevant chunks from ChromaDB
   • Produces 3 anime recommendations

5. Frontend Interface (app.py):

   • Simple Streamlit UI for entering queries
   • Returns recommendations instantly

---

Deployment Guide (on GCP using Kubernetes)

Infrastructure Setup

1. Create GCP VM Instance

   • Install: Docker → Minikube → kubectl
   • Setup: git clone your repo into VM

2. Docker & Minikube

   docker build -t anime-recommender .
   minikube start

3. Secrets

   • Store .env securely
   • Create Kubernetes Secret from .env file

4. Deploy on K8s

   kubectl apply -f llmops-k8s.yaml
   minikube tunnel  # Run in a separate terminal to expose LoadBalancer

Service Types Explained

Type	Use Case	Access
ClusterIP	Internal microservices only	Not exposed outside cluster
NodePort	Expose on VM IP + specific port	http://<VM-IP>:<NodePort>
LoadBalancer	Internet-facing deployment	GCP auto-assigns external IP address

Production Tip: Use LoadBalancer for real deployments.

---

📊 Monitoring with Grafana

1. Create namespace:

   kubectl create namespace monitoring

2. Install Helm & Grafana:

3. Set up Grafana Cloud, get token, and configure observability

---

Environment Variables (.env)

GROQ_API_KEY=your_groq_api_key_here
MODEL_NAME=groq/llama3-8b-8192

---

Sample Query

Ask questions like:

"Suggest some action-packed anime like Attack on Titan"

"What are the best anime series for beginners?"

---

Tech Stack

• Python, LangChain, GROQ LLM
• HuggingFace, ChromaDB
• Streamlit, Docker, Kubernetes, Minikube
• GCP VM, Grafana, Helm