Project goals
- Preserve Seasons 1-7 canon while generating a six-episode alternate Season 8.
- Carry alternate-timeline decisions and open threads across episodes.
- Compare agentic, retrieval, memory, and baseline conditions reproducibly.
Final project overview
Cosine of Thrones is a multi-agent Retrieval-Augmented Generation system that uses vector search, GraphRAG, Neo4j memory, LangGraph orchestration, and narrative evaluation to generate and critique alternate versions of Game of Thrones Season 8.
Long-form narrative generation is not only a text-generation problem. It is a continuity-management, evidence-retrieval, planning, revision, and evaluation problem.
System architecture
Hybrid vector and GraphRAG retrieval ground the story in canon. A season planner, episode planner, storyboarder, writer, editor, patcher, memory writer, and season checker make the generation process visible.
Evaluation design
NSS scores narrative structure and canon fit. Season-level checking, quote-backed consistency evidence, arc-closure coverage, runtime, and token use expose quality and reliability tradeoffs.
Current finding
The latest agentic comparison run outperforms the Redux human fan rewrite under the shared NSS rerun, but it does not outperform original HBO Season 8. The stronger result is inspectability and control.
What GraphRAG adds
GraphRAG creates structured canon pressure through entities, relationships, text units, and community reports. It makes grounding more explainable, while sometimes penalizing plausible creative additions.
Important limitations
LLM-based evaluators can vary, full-season closure remains difficult, and multi-agent planning, editing, repair, retrieval, and scoring require significant runtime and token cost.