Fable

Inspiration

The Korean webnovel "Omniscient Reader's Viewpoint" was the inspiration for this project. I was enthralled by the immersive detail of the world created in the novel, and also fascinated by the power that storied could have, even in a fantastical world.

What it does

Fable helps creators build and explore living worlds.
Users can write lore notes, extract entities/relations/events, visualize everything as a graph, place events on a timeline, and chat with an in-world Historian NPC that answers both factual and speculative questions while staying grounded in canon.

How we built it

Frontend: React + TypeScript for graph/timeline/notes/Historian interfaces.
Backend: FastAPI + SQLite for world state, entities, relations, notes, and timeline operations.
AI integration: Backboard assistant threads for analysis and Historian chat.

Challenges we ran into

Timeline integration: keeping entity/relation state consistent across marker operations and world-state reconstruction.
Handling large data: processing long notes reliably required chunking, retry logic, and timeout-safe flows.
RAG document constraints: fitting rich world context into a limited document budget while keeping retrieval quality high.
Smart context retrieval: getting the right context to the model without overloading prompts or causing answer drift.

Accomplishments that we're proud of

A scalable RAG compiler strategy that organizes whole-world lore into structured document slots.
In-world Historian responses that feel conversational (not generic markdown dumps).
Integrated graph + timeline + notes workflow that turns unstructured writing into queryable canon.
Canon-aware assistant behavior that supports both factual QA and plausible “what-if” reasoning.

What we learned

Context retrieval is very difficult, and a very large, open area of improvement for AI models.
Reliability is absolutely critical for long-term workflows, ensuring that tiny early data misalignments don't ruin large projects hours into building.