# Dense results An agent that retrieves eight passages has to read eight passages, decide which are on-topic, and do the synthesis itself — every time anyone asks. Brain consolidates the sources behind a subject into one summary with a citation per claim, so the answer arrives already made. This page is the same query against the same index, with and without that layer. ## The query *"cyclooxygenase inflammation cancer"* ## Without the summary layer — eight fragments Ranked document hits. Each is a real source; none is an answer. ``` 1 Cyclooxygenase-Dependent Tumor Growth through Evasion of Immunity 2 Tumor-associated macrophages and high COX-2 → poor prognosis 3 COX-2 inhibitors enable repeated vaccinia virus administration 4 PI3Ks qualify as drug targets for cancer, inflammation … (loosely related) 5 Osteoarthritic cartilage degradation indicators … (loosely related) 6 15-Hydroxyprostaglandin dehydrogenase down-regulated in colorectal cancer 7 Omega-3-derived lipid mediators 8 COPD, neutrophils, cigarette smoke and COX ``` Thousands of tokens, two of them off the point, and the synthesis is still the model's job — repeated on every query that lands near this subject. ### The raw response A live pull for `"interleukin-1b inflammasome"` on the same corpus — the rows a plain vector store returns, documents and facts, no synthesis: ```json [ { "id": "700b960dd0a3aad0", "kind": "doc", "score": 0.992, "title": "Human Monocytes Engage Alternative Inflammasome Pathway" }, { "id": "df399593abc4edcb", "kind": "doc", "score": 0.989, "title": "Inflammasome-induced IL-1β secretion in microglia" }, { "id": "4e6192a75e2477bc", "kind": "fact", "score": 0.939, "title": "[fact] Inflammasomes are multiprotein complexes that li…" }, { "id": "86ac2fbf338b865b", "kind": "fact", "score": 0.935, "title": "[fact] Cytokines downstream of inflammasome activation,…" } ] ``` Four rows to read and reconcile before answering anything. ## With the summary layer — one consolidated answer The top result is a single entity summary, consolidated from **11 source documents**, each claim carrying the source it came from: ``` Cyclooxygenase-2 (Cox-2) — an enzyme implicated in inflammatory and pathological processes: viral clearance, lipid-mediator generation, and cancer development. - Viral clearance: Cox-2 inhibitors circumvent viral-clearance limits via neutralising antibodies [cited] - Lipid mediators: generates bioactive signals (15-epi-LX, 18R-HEPE) via NSAIDs + omega-3 [cited] - Renal vasoconstriction [cited] - Cancer: survivin correlation [cited] - COPD airways [cited] ``` The agent cites this and stops. It also carries a small entity-relation graph — cox-2 → survivin, PGE-2, arachidonic-acid, aspirin — so the next question about a neighbouring subject starts from the relations rather than another eight-document read. ### The raw response The same live pull as above, unfiltered. Row 1 is the dense summary, above every raw document and fact — and its `kind` says what it is, so a model can prefer it deliberately rather than by luck: ```json [ { "id": "0a013be2e4b93400", "kind": "summary", "score": 0.997, "title": "[rule] interleukin-1b is a proinflammatory cytokine who…" }, { "id": "700b960dd0a3aad0", "kind": "doc", "score": 0.992, "title": "Human Monocytes Engage Alternative Inflammasome Pathway" }, { "id": "df399593abc4edcb", "kind": "doc", "score": 0.989, "title": "Inflammasome-induced IL-1β secretion in microglia" }, { "id": "4e6192a75e2477bc", "kind": "fact", "score": 0.939, "title": "[fact] Inflammasomes are multiprotein complexes that li…" }, { "id": "86ac2fbf338b865b", "kind": "fact", "score": 0.935, "title": "[fact] Cytokines downstream of inflammasome activation,…" } ] ``` Row order tells the story: the answer, then the primary sources it was built from, then supporting facts. ### The raw document Document `0a013be2e4b93400` — row 1 in full, consolidated from 7 sources: ``` interleukin-1b is a proinflammatory cytokine whose bioactivity is controlled by activation of the inflammasome. Major capabilities and subsystems - proinflammatory cytokine, bioactivity controlled by inflammasome activation [700b960dd0a3aad0] - generated through NLRP3 inflammasome activation in Myelodysplastic Syndromes, driving clonal expansion and pyroptotic death [1a48fcf7bb8cb4be] - produced in human monocytes in response to LPS via an alternative inflammasome pathway [700b960dd0a3aad0] - controlled in ILC2s in airway inflammation, interacting with IL-4 and IL-12 to steer cell fate [b82d5fcd36bb8472, 90250f9be040c89e] Invariants / gotchas - the LPS "alternative inflammasome" uses NLRP3-ASC-caspase-1 but is devoid of pyroptosome formation, pyroptosis and K⁺-efflux dependency [df399593abc4edcb] - microglial IL-1β secretion is only partially caspase-1/4/5-dependent vs hematopoietic macrophages [df399593abc4edcb] ``` The "invariants / gotchas" section carries the exceptions — which is where a synthesis usually goes wrong. ### Relations Every entity carries a typed relations graph, queryable on its own — `t-cells` has 85 edges, `breast-cancer` 57. *Some* summaries, this one included, also inline a machine-readable block of them: ```json { "relations": [ { "target": "caspase-1", "type": "part_of", "confidence": 0.9 }, { "target": "lipopolysaccharide", "type": "integrates_with", "confidence": 0.9 } ] } ``` ## What was measured On BEIR SciFact, a public 5,183-document biomedical benchmark corpus: - **Entity summaries are 35% of top-10 results**, and the **rank-1 result for 25% of queries** — the dense answer surfaces, rather than sitting below the fragments it replaces. - **nDCG@10 of 0.71**, against a BM25 baseline of 0.665 on the same corpus and queries. Consolidation does not cost retrieval quality. The same corpus consolidates `interleukin-1b` from 7 documents and `azacytidine` from 3. Consolidation is per entity, so the ratio follows how much has been written about a subject — and so does quality. Thin input shows: an entity with only a handful of sources can open by defining itself in terms of itself, because there is not enough cross-document material to say something denser. ## Provenance Measured by Levirge's evaluation harness on BEIR SciFact, a public biomedical benchmark corpus, 15 Aug 2026. Both result lists come from the same index and the same query — the only variable is the summary layer. The raw-response sections are a live pull for *"interleukin-1b inflammasome"* on the same corpus, a second query, because it is the one we hold verbatim payloads for. Titles are excerpted; the summary text is verbatim apart from one unrelated line — a consolidation defect we logged rather than quietly leave in the example. The effect is a property of consolidation, not of this corpus: how dense your answers get depends on how many sources say something about the same subject. ## Try it [Request access](/contact), then call `search_knowledge` against your own Vault — the consolidated summaries are what it returns.