plgg-content

UNSTABLE — Experimental study work (POC). Part of the plgg monorepo.

A derived, rebuildable SQLite index over a git-primary Markdown corpus, with an HTTP-agnostic, MicroCMS-like read-only query API and always-on FTS5 (BM25) search, built from scratch on the plgg family (plgg-md, plgg-sql, plgg-db-migration).

The governing decision is D4: content is git/filesystem-primary; SQLite is a derived, rebuildable index. Every row is reconstructable from the Markdown — losing the database costs nothing but a rebuildIndex. The primary operation is therefore a full, idempotent rebuild; incremental indexDocument / removeDocument are optimizations over it.

HTTP-free query core

The query surface is plain typed Db-taking functions, not an HTTP framework — the same functions back the SPA delivery API, the admin UI, the MCP tools, and the Claude Code plugin export. An HTTP layer (e.g. plggpress's /api mount) is a thin adapter over them.

import {
  openIndex,
  registerCollection,
  indexDocument,
  rebuildIndex,
  listCollection,
  getDocument,
  searchIndex,
} from "plgg-content";

// open a ready index (schema created idempotently)
const db = /* Ok */ await openIndex("content.db");

// ingest already-validated pages (ticket-17 content models
// validate upstream; plgg-content stores + serves)
await rebuildIndex(db)(pages);

// MicroCMS-style delivery
await listCollection(db)("blog", listQuery); // { contents, totalCount, limit, offset }
await getDocument(db)("blog", "/blog/hello"); // Option<Document>
await searchIndex(db)("kangaroo", 10);        // BM25-ranked hits, no LLM key needed

Design

• Schema (Schema/) — documents (one row per page, typed frontmatter serialized), chunks (heading-scoped RAG/search granularity), collections (serialized CollectionSchema), and an external-content FTS5 table over chunks with its sync triggers — all emitted through plgg-sql's FTS5 builders (no raw FTS5 SQL). The schema is created idempotently via execScript (D4: the whole DB is rebuildable, so no versioned ledger).
• Ingest (Ingest/) — a pure heading-scoped chunkBlocks fold over plgg-md's Block AST, a transactional idempotent indexDocument (skips an unchanged content_hash), and rebuildIndex (ingest
  • prune vanished paths).
• Query (Query/) — listCollections / listCollection / getDocument / searchIndex, plus the ListQuery caster (bounded limit, closed orderBy, sanitized q) and the serializable CollectionSchema.
• vendors/ — the only driver-aware code: the node:sqlite implementation of plgg-sql's Db seam.

The only runtime dependencies are plgg, plgg-md, plgg-sql, and plgg-db-migration; it is plggpress-agnostic (plggpress depends on it, never the reverse).