Physics / mechanism
GraphRAG denotes retrieval-augmented generation in which the retrieval substrate is a graph of entities and relations rather than a flat vector index. The supplied sources do not describe a GraphRAG pipeline, its indexing procedure or its query-time behaviour; they cover only the storage layer such a system would sit on. What follows is therefore restricted to that layer.
The one substantive source is a hardware-first graph database model called Views, from Serb’s Edinburgh CEF group, published as arXiv preprint 2508.18123, submitted 25 August 2025 with a v2 revision on 13 November 2025 and carrying a header stating it is a preprint, not submitted and not peer-reviewed. Data is stored as chains of linknodes with the field layout [head ID, primID1, prop1, primID2, prop2, next], encoding ternary source-edge-destination relationships in which properties are recursively nestable, so a property can itself carry properties via subordinate chains.
The structural claim against RDF and property graphs is that edges and destination vertices are treated as equivalent entities inside a source-centred structure, and that the linked-list organisation allows traversal to follow next pointers without broadcasting the whole memory. The paper specifies hardware operations rather than a schema alone, including a content-addressable read (CAR) primitive. The relevance to GraphRAG is indirect: cost per traversal step, and whether traversal can be localised in memory, set the economics of graph-structured retrieval at scale. The sources supply no measured figures on that.
Competitive landscape
Evidence base
Frontier (open questions)
- What is the measured traversal latency and energy per linknode hop for the Views model on real hardware, and how does it compare with a conventional property-graph engine on the same query set?
- Has the Views preprint passed peer review, and have any of its ISA operations, including CAR, been implemented in silicon rather than simulated?
- Do graph-structured retrieval pipelines built on such a store show measurable answer-quality or cost gains over vector-index RAG on a published benchmark?
- At what graph size does the claim of traversal without whole-memory broadcast break down, and what is the memory-capacity ceiling per device?
Synthesised 2026-08-31 from 2 KB sources by the resynth pipeline; citations are KB source slugs.