How it works

How it works🔗

FILE.md ──► doc.rs ──────► draft.rs ─────────► scene.rs ──► Bevy
            sections +     WorldManifest       entities,
            BLAKE3 hashes  (localgpt-world-    lights, fog
                            types)                  ▲
   ▲                                                │
watch.rs: poll the file, recompile on save     tour.rs: camera + caption
  • doc.rs: Markdown → title, front matter (genre: …), and one section per place — a ## heading (world) or a ----separated slide (deck). Each section carries a hash of its text (plus any world fence).
  • draft.rs: sections → a WorldManifest with one region per section — a winding path (world) or a straight presentation path (deck) — and a tour with one stop per section. Deterministic, with entity ids that are stable per section, so editing one section leaves the other regions untouched. Per section the first of these wins: a world fence’s exact entities, the sidecar’s cached agent build, the cached recipe — else the rules.
  • agent.rs / assets.rs (llm feature): the tool-calling agent (ported from Verse) and the pack’s manifest/kind/span math. A pure interpreter applies the model’s tool calls to platform-local entities; the pack’s real CC0 models are placed by semantic kind.
  • recipe.rs / sidecar.rs (pure): the clamped per-region recipe type with its lenient LLM-reply parse, and the hash-keyed <doc>.world.json v2 cache (builds + recipes). Compiled in every build; only the authoring is feature-gated.
  • llm.rs / generation.rs (llm feature): the model (ported from Verse) and the worker thread running the tier chain — agent build → recipe → draft — in the background while the draft is already on screen.
  • scene.rs: spawns the manifest in Bevy using the same mapping as LocalGPT Gen, and loads placed GLBs.
  • tour.rs and watch.rs: navigation and hot reload.

Worlds use localgpt-world-types, the format LocalGPT Gen saves as world.ron, and pass Gen’s own validation. --print-ron prints exactly that file.

Develop🔗

cargo test
cargo clippy --all-targets -- -D warnings
cargo fmt --check

LocalGPT MD is open source under Apache-2.0; the code is at github.com/localgpt-app/localgpt-md.