GraphRAG Knowledge Graph Consistency Checker

Construction procurement in regulated markets involves thousands of interdependent clauses, technical specifications, and legal requirements. Manual consistency checking across large tender packages is error-prone and scales poorly.

We developed a Knowledge Graph-powered LLM consistency checker that maps the relationships between requirements, specifications, and submitted documents, then uses GraphRAG to surface inconsistencies and compliance gaps with precise provenance.

We led the full lifecycle from system design through production deployment:

• **Knowledge Graph modeling**: Designed a Neo4j AuraDB schema capturing entities (requirements, specifications, clauses, parties) and their relationships (requires, contradicts, supersedes, references)
• **GraphRAG pipeline**: Rather than flat vector search, queries traverse the graph to surface multi-hop relationships - e.g., identifying when a sub-clause inherited from a referenced standard contradicts a project-specific requirement
• **Ingestion pipeline**: Documents are parsed, entities extracted via NER, embeddings generated, and nodes/edges upserted into Neo4j in a single idempotent pipeline
• **Multi-tenant isolation**: Per-tenant graph namespaces with role-based access control at the Neo4j layer
• **FastAPI deployment**: Async API with streamed graph traversal results and structured inconsistency reports with clause-level citations
• **MLOps practices**: DVC for data versioning, MLflow for embedding model tracking, GitHub Actions CI/CD for zero-downtime deploys