Skip to content

v0.1.8 — 4-Stage Curator Pipeline

Choose a tag to compare

@ssdavidai ssdavidai released this 23 Feb 07:33

What's New

4-Stage Curator Pipeline

Replaced the monolithic single-LLM-call curator with a focused 4-stage pipeline:

  1. Analyze + Write Note (LLM) — Short prompt. Creates one comprehensive note and outputs a JSON entity manifest listing all people, orgs, projects, tasks, etc. mentioned in the source.
  2. Entity Resolution (pure Python) — Deduplicates entities against the vault. Creates new ones with real body content, bypassing templates entirely (fixes placeholder leakage).
  3. Interlink (pure Python) — Wires up wikilinks between the note and all entities, and between entities that reference each other.
  4. Enrich Entities (LLM, per-entity) — For each entity, a focused LLM call appends substantive body content and fills missing frontmatter fields.

Fixes

  • Template placeholder leakage — entities no longer get "Brief description of the project's purpose and goal" as their body
  • Agent cognitive overload — instead of one 2,100-line prompt with 20-35 tool calls, each LLM call is short and focused
  • Thin stub entities — person, org, project records now get real content via per-entity enrichment
  • No enrichment of existing entities — existing records now get updated when new info arrives

Other Changes

  • OpenClaw session clearing and workspace sync across all tool backends (curator, janitor, distiller)
  • Stale OpenClaw lock file cleanup before batch processing

Install / Upgrade

pip install --upgrade alfred-vault

Full Changelog: v0.1.7...v0.1.8