From Notes to Knowledge: Building a System That Actually Scales
Decouple acquisition format from retrieval mechanism. Your notes should respond to how you **think**, not how the tool indexes them.
Most systems fail because they force acquisition and retrieval to occur **synchronously**. The act of typing a note should never block the act of retrieving it later.
If the source disappears (e.g., Slack archive lost), the note still has structural integrity. This forces me to preserve provenance without blocking initial capture.
Enrichment is **idempotent**: re-running it never corrupts a note, only repairs gaps. This lets me change extraction logic later without fear.
Hot plane decay follows a simple power-law: `(access_last_week)**0.85`. This naturally skews toward evergreen notes without static rules.
Cost scales **sublinearly** because storage (R2) and retrieval (Postgres, Redis) are **pay-per-use**, not reserved capacity.
Semantic models **age** faster than you expect. Their training cutoff means terminology drift breaks them before syntax or storage does.
The hallmark of a system that **scales** is that **failures are isolated**: a Redis cache failure never disturbs ingest throughput.