SQLite integration

[anhldbk]

Overview

GBrain is a personal knowledge brain backed by Postgres + pgvector. This work adds SQLite as a zero-infrastructure alternative engine using bun:sqlite (built-in, no dependencies), FTS5 for full-text search, and vec0 for optional vector similarity — all behind the same BrainEngine interface so every command works identically regardless of backend.

What was built:

• SQLiteEngine — full implementation of all 30 BrainEngine methods: pages, chunks, links, tags, timeline, raw data, versions, files, ingest log, config, stats, health, and graph traversal
• SQLite schema — DDL with FTS5 virtual table and three sync triggers (insert/update/delete) to keep full-text search current automatically; WAL mode enabled at connect time
• vec0 integration — loads the native extension at connect, creates a parallel chunks_vec vector table when available, degrades gracefully to keyword-only when not
• gbrain init --sqlite — initializes a local brain at ~/.gbrain/brain.db (or a custom path)
• Engine selection — CLI reads config.engine and routes to SQLite or Postgres transparently
• Shared utilities — extracted validateSlug and contentHash from Postgres engine into src/core/utils.ts, fixing a slug validation regex bug along the way
• File commands fix — gbrain files list/upload/sync/verify previously bypassed the engine entirely via a direct DB connection; refactored to use BrainEngine file methods that work with both backends
• Hybrid search fix — hybridSearch unconditionally called embed() before keyword search; without an OpenAI API key this threw and skipped results entirely; now degrades gracefully to keyword-only
• Docker fallback — Dockerfile.sqlite for environments where the vec0 native extension needs to be pre-installed
• 39 new tests across utils.test.ts, sqlite-engine.test.ts, and fts5-query.test.ts

Disclaimer

I've built with Claude Opus