GraphRAG and knowledge graphs

Cited answers across thousands of documents, using a knowledge graph where plain vector search falls short.

Retrieval pipeline: ingestion, chunking, hybrid search, re-ranking, grounded generation with citations
Reference architecture we adapt to each project.

The problem we solve

Vector search finds passages that look similar. It struggles with questions that span documents, depend on relationships, or need an exact clause or code. A knowledge graph of entities and their links, combined with hybrid search and re-ranking, fixes that and makes every answer traceable to its source.

Confident wrong answers

The model fills gaps when retrieval misses the right passage.

Questions across documents

Who supplies what, which clause overrides which: relationships vector search cannot see.

No way to verify

Answers without citations cannot be trusted in legal, finance or support work.

What we build

Knowledge graph construction

  • Ontology and schema design with your domain experts
  • LLM-based entity and relationship extraction
  • Entity resolution and de-duplication
  • Incremental updates as documents change

Retrieval

  • Hybrid search: BM25, vectors and graph traversal
  • Microsoft GraphRAG local and global search, LightRAG
  • Text-to-Cypher for structured questions
  • Cross-encoder re-ranking

Grounding

  • Citations to page, clause or node
  • Refusal when the sources do not support an answer
  • Permission-aware retrieval per user and team
  • Multilingual corpora

Evaluation

  • Golden question sets from real usage
  • Faithfulness, context recall and answer correctness scoring
  • Plain RAG versus GraphRAG comparisons on your data
  • Regression checks on every change
In practice

What this looks like when we build it

All work and demos

The stack

Neo4jPostgreSQLMilvusLangChainPython
Graph databasesNeo4j, Memgraph, Amazon Neptune
Vector storespgvector, Qdrant, Weaviate, Milvus, Neo4j vector index
FrameworksMicrosoft GraphRAG, LightRAG, LlamaIndex property graphs, LangChain
ParsingLayout-aware PDF and Office parsing, OCR for scans
EvaluationRAGAS, DeepEval, custom LLM-as-judge rubrics

Go deeper on the tools:

How we deliver it

Retrieval assessment

2 weeks. Your documents, a golden question set, a measured baseline and a recommended design.

Production build

4 to 8 weeks. Graph, hybrid retrieval, cited answers, access control and an admin view.

Run and improve

Monthly. New sources, ontology updates, eval tracking and tuning.

Engagement models and pricing

Questions

Do we need GraphRAG, or is normal RAG enough?

Often hybrid search plus re-ranking is enough, and we will tell you so. GraphRAG earns its cost when questions span many documents or depend on relationships between entities. We measure both on your data before you commit.

Which graph database do you use?

Neo4j in most projects, for its Cypher query language, vector index and tooling. Memgraph or Amazon Neptune when performance or your cloud setup points there.

Can it respect document permissions?

Yes. Access rules are applied at retrieval time, so users only get answers from content they are allowed to see.

Have an AI system to build?

Tell us what it needs to do. We reply within one business day with questions, a rough plan and the next step.

Book a call
  • Free 30-minute architecture call
  • NDA before you share details
  • You own all code and IP
  • Reply within one business day