Questions we get asked

Vector RAG retrieves chunks that look like the question. GraphRAG first builds a graph of entities and relationships, detects communities in it, and retrieves through those structures as well, which is what multi-hop and corpus-wide questions need. The engine described in the case study does both and fuses the channels with reciprocal rank fusion.

No. The graph can live in PostgreSQL with pgvector alongside the embeddings, which keeps one database, one backup and one migration path. Neo4j is a reasonable choice when graph traversal dominates the workload.

With a fixed query set and numbers you can rerun: strict hit rate, entity and phrase recall, latency per query, and indexing time. A benchmark that only reports "it works well" is not a benchmark. See retrieval engineering.

Below roughly tens of millions of vectors, PostgreSQL with pgvector usually wins on operational cost; above that, or when you need specialised index types, a dedicated store earns its place. The comparison matters less than the retrieval strategy on top of it.

The Model Context Protocol lets an editor or agent call your tools without a bespoke client. Exposing retrieval over MCP means the same engine serves a chat UI, a code editor and an internal agent without three integrations.

Each claim is stored with its document, page, quote and confidence at extraction time, and the report is rendered from that graph — so a claim without a page reference cannot be printed. Low-confidence edges are re-checked against the source text first. See the Evidence Brief service.

An Evidence Brief starts at $750 and takes five business days. Engineering work is quoted per project with a fixed deliverable list; both start with a written scope. Get in touch.