SilverBullet

003 Indexed Object Graph

Properties1
tagsadr

2 min read

Context

SilverBullet’s power features — Space Lua/Integrated Query, Linked Mention, the Linked Tasks, Object Graph — all need structured, queryable data. But the source of truth is plain markdown files. Something needs to turn that prose into queryable data without making the markdown any less plain.

Decision

Continuously maintain an Object Index: a structured, queryable graph automatically extracted from the markdown. Every meaningful element — pages, headers, tasks, list items, tags, links — becomes an object (a “row”) with attributes, typed by one or more tags (“tables”), queried via Space Lua/Integrated Query and the index API.

Two principles hold it together:

  • Markdown is the source of truth. Every indexed object has a representation in the markdown; the index is derived, can be flushed, and rebuilt from the files at any time (Space: Reindex). It is a cache, not the canonical store.
  • The index lives in the client. It is built into IndexedDB in the browser (ADR/007 Core Application Logic on the Client, ADR/001 Offline-First PWA) and maintained incrementally.

Consequences

Positive

  • Plain markdown stays plain, yet queryable. Structure is derived, so files remain portable and human-editable; nothing is locked in a database.
  • Rebuildable and disposable. The index is a cache. Corruption or schema changes are fixed through a reindex.
  • One uniform model. Pages, tasks, tags, and relations are all just objects with attributes enabling one query language over everything.
  • A real graph for free. Generalized relations turn backlinks and frontmatter links into uniform edges, enabling backlinks, the Object Graph, renames, and semantic frontmatter (like ADR dependsOn/related).

Negative / trade-offs

  • Every client (re)builds the index. Indexing all files is a cold-start cost per client (see ADR/001 Offline-First PWA), large spaces take longer.

Alternatives considered

  • A separate authored database (Notion/Confluence-style). Rejected: breaks “markdown is the source of truth,” adds lock-in and lossy export — contrary to the whole product thesis.
  • Query raw markdown text (grep-style) at query time. Rejected: too slow and too weak, no typed attributes, no joins/graph, no incremental maintenance.

References

Graph View

Backlinks