Section
AI & LLM Ingestion
How to extract Wheel of Heaven content and context for AI models — llms.txt, the JSON API, curated context endpoints, library primary texts, and RAG patterns.
This section is the canonical reference for using the Wheel of Heaven corpus with AI systems — LLMs, retrieval-augmented generation pipelines, agentic assistants, embeddings indexes, fine-tuning corpora, or just a one-off paste into a chat window.
The project is built to be ingested. Every content page on www.wheelofheaven.world — wiki, timeline, articles, news, library, sources — has a deterministic JSON twin on api.wheelofheaven.world, advertised from the page itself via <link rel="alternate" type="application/json">. That twin is 6.3× smaller than the rendered page, so it is the cheapest way to read a page you already have the URL for. (Chrome and utility pages have no twin — see the twin link contract.) Curated narrative summaries are published at stable URLs for direct system-prompt use. The full corpus is available as a single llms-full.txt file. Everything is CC0-1.0 public domain — no auth, no rate limits, no licensing friction.
Start here#
| If you want to… | Read |
|---|---|
| Get an LLM caught up on the project in one paste | Quickstart |
Understand the llms.txt and llms-full.txt manifests | llms.txt and llms-full.txt |
| Discover the machine surfaces automatically, or read the crawler policy | Agent discovery and crawler policy |
| Hit the JSON API from agent code | API endpoints for AI agents |
Use the curated /v1/context/* narrative endpoints | Curated context endpoints |
| Pull primary texts and bibliographic records | Library and sources |
| Build a RAG corpus or embeddings index | Embeddings and RAG |
| Write a good system prompt that grounds the model | System-prompt patterns |
| Credit the corpus correctly | Attribution and licensing |
What makes the corpus AI-ingestible#
The project is built around a small set of design commitments that matter specifically when an LLM is the reader:
- Every page has a JSON twin. The reading site at
www.and the API atapi.are independent Zola builds of the same canonical data. The API isn’t an afterthought — it’s a first-class surface. - One canonical URL per fact. The API guarantees URL permanence for
/v1/. A path you ingest today resolves to the same kind of object next year. (See API changelog.) - Curated narrative endpoints for direct ingestion.
/v1/context/*serves the project’s hypothesis, terminology, timeline, source program, and editorial method as plain prose — designed for paste into a system prompt with no transformation. - Every claim labelled by epistemic status. Pages carry
claim_type = "direct" | "framework" | "inferred" | "speculative", so an ingesting model can be told to surface that label alongside any answer it gives.frameworkis the load-bearing one for an ingesting model: it marks a claim the project’s canon states plainly but mainstream scholarship does not endorse. - A controlled vocabulary for terminology. The terminology endpoint includes a do-not-use table — the project cares which words the model picks (Elohim vs. “aliens”, Yahweh vs. “God”).
- JSON Schemas and enums published. Every response kind has a schema at
/v1/schema/{kind}/. Every controlled vocabulary has an enum at/v1/enums/{name}/. A model can validate its own output against these. - CC0-1.0, no auth, no rate limits. Ingest freely. Attribute when convenient. See Attribution.
- Multilingual mirror. Same surface under
/v1/{lang}/...for 9 languages, with hreflang correctly wired on the reading site.
What this section is not#
- Not API reference. For tabular endpoint catalogues, the response envelope shape, controlled-vocabulary values, and JSON Schemas, see API Reference. This section is task-oriented: what to do, in what order, with what tradeoffs.
- Not a tutorial on LLMs in general. Familiarity with system prompts, tool use, retrieval, embeddings, and basic HTTP is assumed.
- Not opinionated about which model to use. Every pattern here works with any modern LLM. Provider-specific guidance is called out inline where it matters (e.g. context-window sizing).
A note on the project’s epistemic posture#
When you ingest the Wheel of Heaven corpus, you are ingesting a working hypothesis, not a creed. The project itself labels every page’s main claim as direct (what a source asserts and the mainstream accepts), framework (what the canon states plainly and the mainstream does not endorse), inferred (what scholarship reasonably concludes), or speculative (what the project proposes as interpretive synthesis).
If your downstream application surfaces Wheel of Heaven content to end users, the editorial method endpoint at /v1/context/method/ explains how to preserve that discipline. The System-prompt patterns page shows how to encode it for the model.
- Quickstart The fastest path from zero to an LLM that can answer accurately about the Wheel of Heaven project.
- llms.txt and llms-full.txt The two top-level manifests the project publishes for LLM ingestion — what each contains, when to use which, and how they're maintained.
- Agent discovery and crawler policy The machine-facing discovery surface — .well-known endpoints, the API catalog, the MCP server card, the skills index — plus the project's crawler policy and what it deliberately does not implement.
- API endpoints for AI agents Which api.wheelofheaven.world endpoints are best for AI ingestion — the AI-relevant subset of the surface, with payload shapes and example calls.
- Curated context endpoints What each /v1/context/* endpoint contains, when to ingest it, and how to compose them into a system prompt.
- Library and sources How to ingest the primary texts (library books at chapter and verse) and the bibliographic backbone (sources, tradition hubs) for AI use.
- Embeddings and RAG How to build a retrieval-augmented generation corpus from Wheel of Heaven content — chunking strategies, metadata, the search index, and pipeline patterns.
- System-prompt patterns Working system prompts and prompt structures for grounding an LLM on Wheel of Heaven content — claim-type discipline, citation discipline, refusal patterns.
- Attribution and licensing CC0-1.0 licensing terms, suggested citation formats, and how to attribute the corpus in AI-generated output.