Skip to content

Repo graph ingest experiment

repo-graph-ingest is an experimental, file-only way to use an upstream repository knowledge graph as candidate Pack/Observe context.

It is designed for comparison and onboarding workflows:

Understand-Anything knowledge-graph.json
→ compact ContextPack-style artifact
→ lexical baseline vs graph-neighborhood comparison
→ report.md + comparison.json receipts

Safety posture

This experiment does not:

  • install an upstream agent skill
  • write durable memory Store records
  • modify Gateway, plugin, cron, or runtime config
  • treat the upstream graph as source of truth

The graph output is useful as context and navigation support. It should be verified before driving product or ops decisions.

Run the experiment

python scripts/repo_graph_neighborhood_experiment.py \
  --graph /path/to/.understand-anything/knowledge-graph.json \
  --repo-root /path/to/repo \
  --out-dir /tmp/repo-graph-experiment \
  --query "retrieval backend selection" \
  --query "graph context pack creation"

The output directory contains:

  • pack.json — compact ContextPack-style candidate context
  • comparison.json — machine-readable baseline vs graph-neighborhood comparison
  • report.md — human-readable summary
  • manifest.json — graph digest, repo commit, and output map

Metrics

For each query, the experiment records:

  • baseline lexical path count
  • graph path count including neighbors
  • overlapping paths
  • novel graph paths
  • neighborhood novelty rate

These are directional diagnostics, not benchmark claims. Use them to decide whether graph-neighborhood expansion improves a specific Pack-lane workflow before integrating it into a larger pipeline.