I tried "The Wise Man's Fear" and the results were not so great[1]. It got a good number of the characters, but only showed them being related to Kvothe. Harry potter worked better'ish, but still not great[2].

My primary interest would be using these relational graphs to generate better structured summaries of stories in a structured JSON object for improved search/recommendations.

- [1]https://austen.pages.dev/ecbc2d49-c10d-4e49-ac68-0d37a819b52...

- [2]https://austen.pages.dev/ace16c1d-c74e-4737-a98d-5401047a1cd...

The system prompt can be found here:

https://github.com/herol3oy/austen/blob/main/src/server/rout...

It would be more interesting if the graph could be generated by sourcing an arbitrary epub file into a RAG and have it figure out character relations without reading the whole book.

I wouldn't be surprised if it all could be done with a script using aichat[1] and rendering the diagram using graphviz.

[1] https://github.com/sigoden/aichat

For anyone vibe checking the results:

https://novelchallenge.github.io/

At best you’re looking at 60% or so accuracy

I did Dungeon Crawler Carl - hilariously wrong

graph TD A[Carl] -->|Pet| B[Princess Donut] A -->|Enemy| C[Mordecai] A -->|Allies| D[Ellie] A -->|Allies| E[Katia] A -->|Enemy| F[The System] B -->|Rival| G[Bea] D -->|Friend| E E -->|Friend| A

Hi, I created the app. I appreciate all of your comments. I know some may find the result of a book not concise or accurate. One issue was that I couldn't afford the price of a better LM service. At first, I tried a free tier of Gemini, but its results were disappointing. Later, I switched to DeepSeek and got better results.