Transforms

A transform takes a fragment of a mimetype it claims and emits child fragments with typed relations (design/indexing.md). Each transform is a plugin; the registry they sign up with is the transforms seam (../architecture/kernel.md).

One registration pathway, two tiers

Linked (Rust) transforms and loaded (WASM) transforms register the same way: claims(mimetype, is_root) + apply(ctx) -> output, plus a per-run LLM call budget and a shape fingerprint (a config digest; the artifact version for loaded ones) that feeds the shape stamp (maintenance.md). Registration returns an undo function held as a fiber effect — unmounting the plugin runs it, and the next sweep notices the difference.

The sweep applies transforms recursively: over the source root, then over every fragment they emit, until nothing claims the output. A transform claiming another’s emitted mimetype (the loaded tier’s usual shape — e.g. OCR emitting text/plain from images) chains in the same rebuild. Application order is structural transforms before enrichment ones, then by entry id — deterministic no matter what order plugins started in.

Capabilities are handed in, never grabbed: the context carries the text, optionally the raw bytes (only for transforms that declare wants_bytes, and only at the root), and optionally a granted LLM (GrantedLlm) — metered against the transform’s per-run budget and checked against the LlmCall guard on every call. Budget spent, or guard says no → the LLM handle refuses, and the transform falls back gracefully.

Output is uniform: sprouts (child fragments with relation kinds — a summary is a sprout related derived-from) plus entities, which the sweep deduplicates across the index and wires mentions edges for, since a transform can’t know fragment ids.

The first-party transforms

• Markdown (structural, text/markdown roots) — splits the document by its own heading outline; sections nest as written, http(s) links become text/uri-list children (links-to). Documents with no headings fall back to chunking.
• Chunker (structural, other indexable text roots) — paragraph-boundary chunks aimed at target_chars (hard break at twice that), with line extents.
• Summarizer (enrichment, every root) — mandatory: every indexed source gets a summary fragment (text/x-inseam-summary, related derived-from the root). Three qualities depending on circumstance: LLM (when the capability is granted), extractive (when there’s text but no LLM), envelope-derived (when the content was never read). The provenance rides the mimetype as via=llm|extractive|envelope.
• Entity extractor (enrichment, every root) — uses the LLM to pull out up to max_per_source named entities; each becomes one text/x-inseam-entity;kind=… fragment per index, with mentions edges from every fragment whose text contains the name.
• OCR (plugins/ocr, loaded, image roots) — the reference community-tier transform: reads image text through the granted vision LLM, emitting text/plain;via=ocr related transcribes (../plugins/loaded.md).

The sweep entry’s limits prune every transform’s output: a depth cap and a per-source fragment cap.

Mimetypes the index defines

Mimetype	Meaning
text/x-inseam-summary;via=<provenance>	the mandatory summary fragment
text/x-inseam-entity;kind=<kind>	a deduplicated entity
text/uri-list	a link found inside a fragment

These are inseam’s own derived types: no transform may claim them, and the wasm bridge refuses components that try to emit them.