CSC 675 · Final Project · Carlos Ortiz
From GraphRAG to Writers' Room
An inspectable multi-agent system that retrieves canon, plans an alternate Season 8, writes and revises episodes, carries story memory, and evaluates the result.
average NSS / 160
episode treatments
and project artifacts
the completed ablation
The agent roles
The published work is more than a final script. Each stage leaves behind inspectable decisions, evidence, critiques, repairs, and evaluation results.
Explore the work
A single front door for professors and students: read the new story, inspect how it was produced, and compare the results against HBO, Redux, and ablation baselines.
Read the New Alternate Season 8
Open the visual plot board, episode treatments, generated screenplay, revised screenplay variants, and scene images.
Evaluation & Artifact Library
Browse NSS results, raw JSON, comparison summaries, source treatments, and ablation evidence.
Generated Episode 6 Script
Read the generated finale screenplay and inspect the associated visual scene artifacts.
Final project overview
The overview connects the implementation to the paper’s argument: long-form generation is a continuity, retrieval, planning, revision, and evaluation problem, not only a text-generation problem.
Core claim
Agentic RAG makes story decisions explicit, grounded, reviewable, and repairable across a full season.
Architecture
Vector retrieval, GraphRAG, Neo4j alternate memory, and LangGraph orchestration work as one pipeline.
Evaluation
NSS, season-resolution checks, consistency evidence, arc closure, runtime, and cost expose different tradeoffs.
Finding
The agentic run beats Redux in the shared rerun, but HBO remains highest and finale closure remains difficult.
Limitations
LLM evaluators vary, GraphRAG can constrain creativity, and multi-agent quality carries substantial token cost.