Renovate Graph RAG inference engine

[nolanefe]

Summary

Renovates the Graph RAG inference engine into a schema-aware, provenance-aware RAG system with deterministic Cypher validation and reproducible evaluation.

Highlights

• Added typed planner, retrieval, evidence, and grounded-answer contracts
• Added schema-aware constrained Cypher planning
• Added deterministic read-only Cypher validation
• Added directional relationship validation across chained graph patterns
• Added rejection of unsupported variable-length relationships
• Added Mock and OpenAI model providers
• Added Fake and Neo4j graph stores
• Added provenance-aware evidence construction
• Added bounded synthesis context with retrieval semantics
• Added structured grounded-answer synthesis and citation validation
• Added reproducible offline and live planner evaluation
• Added real Neo4j AuraDB integration validation
• Added benchmark/result documentation and portfolio-focused README

Evaluation

• 200 offline tests passing
• Ruff clean
• mypy clean
• Offline evaluation: 10/10 pipeline cases, 19/19 safety cases
• Controlled 60-request GPT-4o-mini planner comparison documented in repository
• Final real Neo4j traces cover direct, multi-evidence, multi-hop, and absent-evidence cases

Validation

• python -m pytest
• ruff check .
• mypy src
• git diff --check

No live credentials or secrets are committed.