name: dirac-vector-kb-ops
description: Use when the user asks to build, refresh, validate, or inspect the Dirac DFT/TDDFT vector knowledge base, including ingestion status and retrieval quality.
Dirac Vector KB Ops Skill
Use When
- User asks to "supplement vector KB", "refresh KB", or "check KB sync".
- Need deterministic proof that KB ingestion happened and sync state was updated.
- Need a quick retrieval probe after ingestion.
Primary Tool
scripts/run_vector_kb_ops.py
Typical Commands
# Full workflow: ingest + query probe + sync snapshot
python scripts/run_vector_kb_ops.py --mode full --base-url http://127.0.0.1:8001
# Force reingest and include web references from manifest
python scripts/run_vector_kb_ops.py --mode ingest --force-reingest --include-web
# Query-only probe
python scripts/run_vector_kb_ops.py --mode query --query "Octopus TDDFT convergence" --top-k 5
# Status-only snapshot from sync file
python scripts/run_vector_kb_ops.py --mode status
Required Outputs
- Ingestion command exit code.
- Ingestion report path from
kb_sync.report_json / kb_sync.report_md.
- Query endpoint used and
hits_count.
- Sync snapshot fields:
total_sources, ingested_sources, total_chunks_added.
Rules
- Always return structured evidence from tool output before narrative summary.
- If ingestion fails, report stderr and stop claiming KB is updated.
- For incremental no-change runs, explicitly report
unchanged_sources.
- If retrieval hits are empty, suggest concrete source/topic gaps.
By