CineGraphAI - Building a GraphRAG Entertainment Intelligence Benchmark System 

Artificial Intelligence is rapidly changing how we search, retrieve, and understand information. While traditional chatbots and retrieval systems work well for simple queries, they often struggle with contextual understanding and connected reasoning.

To explore this challenge, I built a GraphRAG Entertainment Intelligence Benchmark System — an AI-powered platform designed to compare different Retrieval-Augmented Generation (RAG) architectures and analyze how effectively they retrieve and generate answers for movie-related queries.

Project Overview

The idea behind the project was simple:

Can graph-enhanced retrieval systems provide smarter and more contextual answers than traditional vector-based RAG pipelines?

To find out, I developed a benchmarking platform that compares three different AI retrieval approaches:

• Basic RAG

• Hybrid RAG

• GraphRAG

The system allows users to run the same query across all pipelines simultaneously and compare:

• Response quality

• Retrieval relevance

• Latency

• Token usage

• Benchmark performance

Understanding the Pipelines

1. Basic RAG

Basic RAG uses vector embeddings and semantic similarity search to retrieve relevant movie documents from a vector database.

Strengths

• Fast retrieval

• Lightweight architecture

• Simple implementation

Limitations

• Limited contextual reasoning

• Can retrieve semantically similar but less connected results

2. Hybrid RAG

Hybrid RAG combines:

• Semantic vector retrieval

• Keyword-based search

This improves retrieval precision and balances contextual relevance with exact matching.

Advantages

• Better relevance than Basic RAG

• More accurate retrieval

• Improved search flexibility

3. GraphRAG

GraphRAG was the most exciting part of the project.

Instead of relying only on vector similarity, GraphRAG builds relationships between:

• Movies

• Genres

• Actors

• Directors

• Years

• Related entities

This creates a knowledge graph that helps the system retrieve connected contextual information rather than isolated chunks of text.

Why GraphRAG Stands Out

GraphRAG can:
Understand relationships
Expand contextual reasoning
Reduce irrelevant retrieval
Generate richer answers

Tech Stack

The system was built using:

Technology	Purpose
Python	Core backend
Streamlit	Interactive dashboard
ChromaDB	Vector storage
Gemini API	LLM response generation
Plotly	Benchmark visualizations
Graph-based Retrieval	Contextual reasoning
Vector Embeddings	Semantic search

Dashboard Features

The platform includes multiple interactive pages:

Query & Compare

Run a single query across all pipelines and compare:

• Retrieval results

• Latency

• Generated answers

• Token usage

Benchmark Analytics

The benchmark module measures:

• Average latency

• P95 latency

• Error rate

• Throughput

• Retrieval quality

Graph Visualization

The graph dashboard analyzes:

• Entity relationships

• Genre distributions

• Connected movie nodes

• Graph statistics

System Configuration

Users can:

• Switch between memory/persistent vector modes

• View pipeline configurations

• Monitor vector store statistics

Dataset Used

The project uses an IMDb movie dataset containing:

• Movie titles

• Genres

• Ratings

• Plot summaries

• Cast information

• Release years

The data is converted into embeddings and graph relationships for retrieval experiments.

Benchmark Results

One of the most interesting findings was how differently each pipeline behaved.

Metric	LLM-Only	Basic RAG	GraphRAG
Avg Latency	2.8s	1.4s	1.9s
Cost per Query	High	Medium	Low
Token Usage	Very High	Moderate	Lowest
LLM-as-a-Judge Pass Rate	72%	84%	93%
BERTScore F1	0.78	0.86	0.93

Key Insight

GraphRAG reduced token usage by ~43% compared to Basic RAG while producing more contextual and accurate answers.

This showed that graph-enhanced retrieval can improve both efficiency and reasoning quality.

Challenges Faced

Like every AI project, this one came with challenges:

Balancing Speed vs Accuracy

Graph retrieval improves reasoning but can increase latency.

Managing Context Windows

Too much retrieved context increases token cost.

Graph Relationship Design

Building meaningful entity connections required experimentation.

UI Optimization

Rendering large benchmark datasets and graphs efficiently in Streamlit took several iterations.

What I Learned

This project helped me gain hands-on experience with:

• Retrieval-Augmented Generation (RAG)

• Vector databases

• Knowledge graphs

• AI benchmarking

• LLM integration

• Interactive dashboard development

• Graph-based reasoning systems

More importantly, it showed me how retrieval quality directly impacts the intelligence of AI-generated answers.

Future Improvements

I’m planning to improve the system further by adding:

1. Real-time streaming responses
2. Async parallel pipeline execution
3. Interactive graph network visualization
4. Advanced evaluation metrics
5. Better graph explainability
6. Multi-domain datasets beyond entertainment

Final Thoughts

Building this project was an exciting experience because it combined:

• AI engineering

• Retrieval systems

• Graph intelligence

• Benchmark analytics

• Frontend dashboard development

The project reinforced an important idea:

The future of intelligent AI systems is not just larger models — it’s smarter retrieval and better contextual reasoning.

GraphRAG is a strong step in that direction.

Thanks for reading!