GraphRAG Consulting

GraphRAG consulting is the design and build of a permissioned knowledge graph used at retrieval time, alongside hybrid chunk search, so multi-hop questions can join entities across documents. Answers still cite spans and refuse when traversal or retrieval is empty. We do this in the client cloud, you own the IP, and a scoped slice is four weeks. It is not a graph for its own sake.

When questions need a join the chunk window cannot see: who owns the vendor that issued the policy that covers this SKU. If hybrid RAG (BM25 plus embeddings) already hits, a graph adds extraction error and ops cost. We measure the chunk baseline first. GraphRAG is justified by golden multi-hop misses, not by architecture fashion.

We implement the pattern — entities, relations, community or neighborhood retrieval, then generation — in your cloud, model-agnostic, with query-time ACL and CI evals. We do not require a specific open-source recipe. If a published GraphRAG pipeline cannot enforce per-user ACL or run in the client VPC, we will not drop it in as-is.

Nodes and edges inherit ACL from source documents. Traversal is filtered to the asking user at query time. A path that crosses a document the user cannot open is cut. Empty authorized traversal refuses. Building one global graph with a service account and filtering in the UI is a leak.

A scoped domain — one ontology, one corpus, chunk baseline, graph path, evals — is four weeks standard, then 30 days on-call. Broader enterprises take multiple slices. Financial services typically 8–12 weeks. Healthcare typically 10–14. Extraction quality, not graph software, is what slips the calendar.

Yes. The graph finds candidate entities and relations. BM25 plus embeddings still fetch the prose those nodes came from. Identifiers and local policy text remain lexical problems. Graph-only generation without cited spans is how GraphRAG hallucinates with confidence.

Your domain owners, after handover. We leave correction queues, extraction evals, and runbooks. Client owns IP. A graph that only the consultancy can edit is not a production system. Thirty days on-call covers extraction incidents; ontology expansion after that is a new fixed scope.

Fixed-scope implementation fee plus platform fee, quoted in writing. You pay inference for extraction and answering. No shared training. Work in the client cloud. Model-agnostic. If the diagnosis is that chunk RAG is enough, we stop at that diagnosis rather than selling a graph.