Content engine

The content engine compiles file-based Assets into a portable query database. JSON and Markdown documents can reference other Assets without embedding their complete source in that database.

Document references

A reference is an object with exactly one $ref property whose value is a URI reference:

{
  "author": { "$ref": "./authors/ada.json" }
}

References can appear at any depth in a JSON document. In Markdown, references can appear at any depth in YAML frontmatter:

---
title: Building with Markdown
author:
  $ref: ./authors/ada.json
---

Article body.

An object with additional properties is ordinary content, not a reference. The $ref spelling is familiar from JSON Schema, but this syntax does not implement JSON Schema resolution, identifiers, anchors, or schema evaluation.

Reference paths use standard URI-reference resolution relative to the source document. The fragment selects which representation of the target to insert:

| Reference | Selected target value | | ----------------------------- | ----------------------------------------------- | | ./author.json | Complete JSON value | | ./author.json#/profile/name | Value at the JSON Pointer /profile/name | | ./article.md | Original Markdown source, including frontmatter | | ./article.md#body | Markdown body without frontmatter | | ./article.md#frontmatter | Markdown frontmatter as an object |

JSON Pointer escaping follows RFC 6901. For example, #/a~1b selects the property named a/b. JSON Pointer fragments apply only to JSON documents; #body and #frontmatter apply only to Markdown documents.

Authored URLs identify Assets during compilation. They do not permit arbitrary runtime network requests. Every target must match a document in the compiled Asset catalog.

Resolution behavior

At build time, the content engine:

1. Parses bounded JSON sources and Markdown frontmatter.
2. Resolves each static URI reference to a document identity and revision.
3. Rejects missing targets, duplicate identities, revision conflicts, and dependency cycles.
4. Stores the validated dependency graph with searchable metadata while source payloads remain in Asset storage.

At runtime, the server compares the query's filter, sort, and output property paths with the graph's reference locations. If a reference can affect the query, the server resolves the relevant source documents before executing the query; otherwise it uses the normal query path without loading graph sources. Non-reference filters first narrow the candidate sources. The server follows only root references that intersect the query, preserves unselected sibling markers, and computes the selected references' complete downstream dependency closure. It loads independent documents concurrently, loads shared storage content once, validates storage identities, revisions, formats, and payload limits, and assembles dependencies before their consumers. A resolved value replaces its exact { "$ref": "..." } marker.

Filtering, sorting, pagination, and projection therefore operate on resolved values when their property paths intersect a reference. Only properties selected by the Assets query are returned. Resolution cannot expose an unselected property merely because it exists in a fetched document. markdown-body-ref content selection is also graph-backed. Compilation retains the Markdown document identity instead of its source bytes, and runtime hydration fetches only the documents selected after filtering, sorting, and pagination. The resolved query result continues to expose the body at content.text. Whole-document and range content modes retain their embedded-content behavior.

Document source caches are keyed by storage reference, revision, and format, so content from one identity cannot be combined silently with different graph metadata. Published runtime assets must also point at the graph node's exact storage reference before the source is fetched. Resolution is bounded by document-count, per-document byte, aggregate-byte, and concurrency limits. A missing document, invalid representation, stale identity, limit violation, or cancellation fails the query instead of returning a partially assembled graph.

Assets without a document graph keep the existing embedded-content behavior and do not require migration.