Long-form fiction poses a persistent problem for LLM-driven writing: local fluency is easy, but maintaining consistent settings, character arcs and foreshadowing across dozens of chapters is hard. This project treats novel generation as a multi-stage, stateful pipeline—combining explicit world/character settings, vector retrieval for long-context recall, and a state-tracking system to surface and preserve plot-level constraints.
What Sets It Apart
- Multi-stage generation workflow: separates setting/workshop, chapter outline, draft generation and finalization so each step can be inspected and corrected—so you get repeatable, incremental control over large narratives.
- State & foreshadowing management: stores character trajectories and plot hooks as structured state, then uses them to detect contradictions and keep development arcs consistent—so later chapters reference earlier events reliably.
- Vector-backed context retrieval: integrates embeddings and a local vectorstore to fetch long-term context (settings, finalized chapters, notes) when creating new chapters—so the model can access far more history than token windows allow.
- Practical multi-provider support: GUI and config support local (e.g., Ollama) and cloud LLM/embedding endpoints, letting users pick based on cost, latency or privacy—so it fits both local-first and cloud workflows.
Who It's For
Great fit if you want to prototype or produce long serialized fiction with a controllable, inspectable pipeline: writers who iterate on outlines, hobbyists building multi-chapter stories, and developers experimenting with RAG-style narrative systems. The project has significant community interest (several thousand stars) and a visible dev-branch refactor noted in the README.
Look elsewhere if you need a plug-and-play hosted SaaS (this is a repo you run locally or host yourself), or if you require production-grade uptime and commercial support. Expect to supply LLM/embedding API keys and possibly resolve dependency/build issues for certain platforms.
