silverstein/minutes/minutes-ingest

/minutes-ingest

Process meetings through the knowledge extraction pipeline to update person profiles, append to the knowledge log, and maintain the index.

Prerequisites

The [knowledge] section must be configured in ~/.config/minutes/config.toml:

[knowledge]
enabled = true
path = "/path/to/knowledge/base"
adapter = "wiki"  # or "para", "obsidian"
engine = "none"   # or "agent" for LLM extraction
min_confidence = "strong"

If not configured, explain what's needed and offer to help set it up.

How to run

Single meeting

minutes ingest ~/meetings/2026-04-03-strategy-call.md

All normal meetings (backfill)

minutes ingest --all

Preview without writing (recommended first time)

minutes ingest --all --dry-run

What it does

1. Reads each meeting's YAML frontmatter (decisions, action_items, entities, intents)
2. Extracts structured facts with confidence levels and source provenance
3. Updates person profiles in the knowledge base (adapter-dependent format)
4. Appends to log.md with a timestamped entry for each ingested meeting
5. Skips facts that already exist (deduplication) or are below the confidence threshold
6. Excludes meetings designated sensitivity: restricted from automated knowledge-base ingestion

Safety guarantees

• `engine = "none"` (default): Only extracts from parsed YAML frontmatter. No LLM involved, zero hallucination risk.
• Confidence thresholds: Facts below min_confidence are counted as "skipped" but never written.
• Provenance: Every fact records which meeting it came from and when.
• Deduplication: Facts whose text already appears in a person's profile are skipped.
• Dry-run: Always suggest --dry-run first if the user hasn't used ingest before.

Interpreting the output

Ingesting 73 meeting(s) into knowledge base at /path/to/kb
  2026-04-03-strategy.md — 4 written, 1 skipped — Mat, Dan
  2026-04-05-standup.md — 2 written, 0 skipped — Alice
  SKIP 2026-03-18-test.md: no frontmatter
 
Done. 6 fact(s) written, 1 skipped, 1 error(s), 3 people updated.

• written: facts that passed confidence threshold and didn't already exist
• skipped: facts below confidence threshold (logged, not written)
• SKIP: files that couldn't be parsed (no frontmatter, invalid YAML, etc.)

Gotchas

• Meetings without summarization have no structured data — If a meeting was recorded before summarization was enabled, its frontmatter won't have action_items or decisions. The ingest will correctly extract 0 facts. This is expected, not an error.
• `engine = "agent"` requires an AI CLI — If the user wants richer LLM-based extraction from transcript body text, they need claude, codex, gemini, opencode, or pi on PATH.
• PARA adapter writes `items.json` — If the user's knowledge base uses the PARA format, facts go into areas/people/{slug}/items.json with atomic fact schema (id, status, supersededBy).
• First run should be dry-run — Always suggest minutes ingest --all --dry-run before the first real run so the user can see what would be extracted.