I’m evaluating AI-assisted writing tools for a personal fantasy worldbuilding project. My primary use case is generating consistent lore, character backstories, and descriptive prose fragments to be consumed via a private API I’m building for a custom world management app. I’ve narrowed the contenders to NovelAI and Sudowrite, but from a backend developer’s perspective, I’m analyzing them as two distinct API platforms with different architectural trade-offs.
My technical requirements:
* **Output Consistency:** The model must adhere to detailed, nested context (hierarchical lore bible) without frequent contradiction.
* **Structured Data Potential:** Ability to generate text that can be later parsed into JSON-like structures (e.g., character attributes, location details).
* **Prompt Engineering Depth:** Fine-grained control over style, tone, and content via system-level instructions.
* **Throughput & Cost:** Efficient generation of medium-length text blocks (300-800 tokens) without prohibitive per-call cost.
Initial benchmark observations from my trials:
**NovelAI (Kayra model)**
* Strengths: Unparalleled context window (up to 8k tokens). The “Lorebook” feature acts like a dedicated, injectable context database, which is architecturally similar to a vector store for key entities. This allows for remarkable long-term consistency.
* Weaknesses: The raw prose quality can sometimes feel less “literary” compared to GPT-4 based offerings. Its API is less RESTful and more niche in its parameter set (e.g., “Bad Words IDs”).
**Sudowrite (GPT-4 & Claude integration)**
* Strengths: Superior prose fluency and “understanding” of complex instructions out-of-the-box. Features like “Rewrite” and “Describe” are effectively pre-built API endpoints for specific transformations.
* Weaknesses: Context management is more opaque. It feels like a curated application layer on top of a foundation model, giving less direct control over how your persistent world data is weighted during generation. The pricing model is more oriented towards direct user interaction rather than high-volume, automated API calls.
For my specific integration, I'm leaning towards NovelAI because its Lorebook system provides a deterministic, configurable way to maintain state, which is critical for worldbuilding. It resembles a specialized database layer integrated into the inference pipeline. However, Sudowrite’s output quality for descriptive passages is noticeably higher.
Has anyone conducted a similar integration or performed systematic prompt engineering tests between these platforms for generating structured world data? I’m particularly interested in any attempts to format output as pseudo-JSON or key-value pairs reliably.
benchmark or bust
benchmark or bust
You're focusing on the API but glossing over the data handling. The "Lorebook" is just a feature - where is that context stored and processed? Their privacy policy and data retention for API calls matters more than the window size.
If you're feeding a "hierarchical lore bible" into a third party API, you've already lost control of that data. You're building a custom app but outsourcing the core intelligence to a black box. That's a foundational risk.
Have you evaluated if you can run an open model locally? Even with a smaller context window, you keep the data. Your consistency requirement is a data governance problem first, a model problem second.
Least privilege is not a suggestion.
Finally, someone asking about the data instead of the features. You're right about the privacy policy being the starting point, but even that's a comfort blanket. The real question is what they train on tomorrow, not just what they store today.
> where is that context stored and processed
Exactly. "Lorebook" is a great marketing term for what's just a prompt prefix in their vector database. The moment you hit the API, your world bible is a training data candidate unless explicitly excluded, and even those exclusions have a shelf life.
And yes, the local model suggestion is the obvious retort. But for the original poster's use case of a private API, the latency and hardware cost of running a 70B model with a decent context window might be prohibitive. They're trading a data risk for a complexity and cost problem. Neither option is good.
Data skeptic, not a data cynic.