Graph Maintenance

The graph improves over time, but changes in models and content can create gaps.

What this page covers

How to identify weak graph areas and re-run extraction safely.

Core workflow

1. Find likely extraction gaps:

pgm-admin sql "
SELECT id, char_length(content) AS chars
FROM entities
WHERE type = 'document'
  AND extraction_status = 'completed'
  AND NOT EXISTS (
    SELECT 1 FROM edges WHERE source_id = id AND source = 'llm-extraction'
  )
ORDER BY chars DESC
LIMIT 20
"

2. Targeted refresh:

pgm-admin reextract --type document --no-edges-only

3. Full refresh with cleanup:

pgm-admin reextract --all --clean-edges

4. Quality and edge hygiene:

• prune-edges --below <threshold>
• validate-edges --limit 200

Notes

• reextract clears extraction state but keeps non-LLM edges intact.
• improve-graph can rerun with alternate model/provider per batch.
• Validate before production rollouts; use --dry-run on risky commands.