[P3] GraphRAG Knowledge Graph — relationship graph over accumulated knowledge

[sou350121]

Summary

Convert Git commit history and Entity Tracker index into a relationship graph: paper ↔ author ↔ benchmark ↔ lab ↔ method; supports structured traversal queries.

Rationale

Deferred: Entity Tracker (P2) satisfies most retrieval needs first. GraphRAG's index construction is O(n²) in LLM calls and only becomes cost-effective at 6+ months of accumulated data or with significantly cheaper models than are available today.

Acceptance criteria

• Graph built from entity-index.json (P2 prerequisite) + Git commit history
• Node types: Paper, Author, Lab, Benchmark, Method, Date
• Edge types: authored_by, affiliated_with, evaluates_on, builds_on
• Query: graph-query.py --from "Lab:DeepMind" --rel "evaluates_on" --to "Benchmark:*"
• Storage: lightweight graph DB (e.g. networkx + pickle, or SQLite adjacency list)

Dependencies

• Requires #[Entity Tracker P2 issue] to be complete first
• Recommended: start after 6+ months of Entity Tracker data has accumulated

Notes

• Deferred from P2 due to O(n²) LLM cost for relationship extraction at scale