Global queries and aggregation over a graph

Classic retrieval selects the k most relevant passages and leaves the generator to compose an answer from them. That paradigm assumes the answer is contained in a small number of units. It fails on global queries, whose answer is written nowhere and must be aggregated across the corpus: “which projects used this supplier, and which of them hit the same failure mode?” requires cross-referencing project write-ups, supplier records and incident reports. Raising k does not help: the context window is bounded, and ranking by local relevance guarantees no coverage.

Edge and co-authors proposed in 2024 to index not just the text but a structure extracted from it. A language model traverses the corpus to extract entities and relations, forming a graph; that graph is then partitioned into communities, for which hierarchical summaries are precomputed. A global query is handled by interrogating those summaries and aggregating the partial answers. The authors report clear gains in answer comprehensiveness and diversity over a vector RAG baseline.

DeepView builds this graph at indexing time, inside your environment. Our implementation departs from the original work on one deliberate point: rather than precomputing community summaries and aggregating them per query, whose cost grows with corpus size independently of the question, we anchor the query on the entities it mentions and traverse the graph from those points, pruning branches as relevance decays. This choice favours latency and traceability at the cost of weaker coverage on genuinely global queries, where the summary-based approach remains superior.