Context Graph

A node is anything the graph can reason about. Nodes have a stable ID, a type, canonical metadata, and any embeddings, labels, or classifications Academe has generated for them.

Edges encode the relationships that make the graph useful for reasoning. Academe preserves the edges we can observe directly (citations, authorship) and infers the ones we can measure reliably (topical similarity, method overlap, contradiction).

• Cites / cited-byReference lists are preserved when available. A citation edge records the citing work, the cited work, the section it appears in, and, when we can extract it, the sentence that justifies the citation.
• Authored-byLinks authors to works. When a paper is retracted or corrected, the edge can carry that status into author and project views.
• Co-authored-withAuthor ↔ author edges weighted by paper count and recency. Powers collaboration suggestions and disambiguates authors with common names.
• Affiliated-withAuthor ↔ institution edges, timestamped so a move from MIT to Stanford is a new edge rather than a replaced one.
• Published-inWork ↔ venue, with issue, volume, and page metadata so citations generated by Academe are bibliographically complete.
• Funded-byWork ↔ funder via grant acknowledgement. Useful for tracking research agendas and for compliance reporting.
• References-dataset / implements-methodExplicit links from papers to the datasets they use and the methods they implement. Makes "find me papers that used this dataset" a one-hop query.
• Topical similarityDense semantic similarity between works (and between works and your notes) via a scientific-text embedding. Similarity edges are thresholded so the graph stays sparse.
• Concept assignmentWork ↔ concept edges with confidence scores, so a paper sits in multiple concept neighbourhoods at once.
• Supports / contradictsClaim ↔ claim edges when Academe’s reviewer pipeline detects an agreement or a disagreement between two extracted claims. Drives contradiction surfacing.
• Version-ofPreprint ↔ published version, or v1 ↔ v2 of the same preprint. Your citations stay consistent as versions replace each other.

The global graph is assembled from a curated set of primary sources, fused into a single normalised index, and enriched with Academe’s own semantic and claim layers.

• IngestPublisher metadata feeds, preprint servers, OA full-text corpora, dataset registries, institutional repositories, funder APIs, and retraction databases. Each source runs on its own schedule, such as hourly for preprints, daily for journals, and weekly for patents.
• ReconcileA multi-stage matcher reconciles DOI, title, author, venue, and year across sources so the same paper appearing in multiple places collapses to a single node with its known identifiers attached.
• NormaliseAuthor names are resolved against ORCID, institutions against ROR, venues against ISSN registries, and funders against the Crossref Funder Registry. This reduces duplicate records and string-spelling drift.
• EnrichWe add concept labels, a dense embedding, a full-text segmentation (sections, figures, tables, equations), a retraction status, and the list of other works that co-cite or are-cited-with this one.
• Extract claimsA claim-extraction model reads full text (where the licence permits) and emits atomic claim nodes along with their evidence spans. Claims get their own embedding so they can be retrieved independently of the papers they come from.
• EvaluateRetrieval and extraction releases are checked against public scholarly benchmarks and internal regression questions. We track deltas so regressions are visible before release.

• DeduplicationSame-work collapse across sources using a hybrid deterministic and learned matcher. We track false-merge and false-split rates as part of release quality.
• CanonicalisationRecords have a canonical ID surfaced in the interface, with alternate IDs preserved as redirects where possible. Exported citations keep the best available identifier.
• Retraction-awareRetractions and corrections flow into the graph from retraction sources. Academe flags retracted papers you’ve cited when the signal is available.
• VersionedPreprint-to-published edges are explicit. Your citations can resolve to the current version unless you pin a specific one.
• ProvenanceEnrichments such as concepts, embeddings, claims, and similarities carry model IDs and timestamps. Reproducibility is part of the data model.
• Licence-awareFull-text access is gated by publisher and repository terms. Paywalled full text is only used when your institutional subscription or BYOK credentials grant the right to.
• Bench-testedReleases run against BEIR, SciDocs, TREC-COVID, SciRepEval, plus an internal scholarly QA regression set. Regressions are tracked before release.

• "Find me the papers I should have cited in §3"Anchor at §3 draft node → expand via concept + similarity edges → filter to works not already in your bibliography → rank by reviewer-likelihood → return five candidates with evidence spans.
• "Has anyone contradicted the claim in my abstract?"Anchor at abstract draft node → extract atomic claim → traverse contradicts edges → return contradicting claims with their sources and the strongest counter-evidence.
• "What methods do adjacent fields use for this problem?"Anchor at your methods section → extract method nodes → walk method ↔ work ↔ concept edges into neighbouring concept clusters → rank by transfer likelihood → return method options with representative papers.
• "Who should I collaborate with on follow-up work?"Anchor at your whole-project embedding → walk to author nodes whose work is most similar → filter by affiliation availability → rank by topical overlap and publication recency → return authors with the papers that placed them in the neighbourhood.
• "Summarise what my project says about X"Anchor at concept X → walk inward into your project graph → gather notes, paragraphs, and claims touching X → return a synthesis with links back to each source.

You don’t have to think about the graph. You write, read, cite, and chat; the graph updates in the background and quietly makes the product more useful.

• StorageThe canonical graph lives in a sharded Postgres layout with optimised adjacency tables. Embeddings are stored in a dedicated vector index. Full-text sits in an OpenSearch cluster for BM25 retrieval.
• EmbeddingsWorks, authors, concepts, claims, and your notes are embedded with a scientific-text encoder tuned on scholarly QA pairs. Re-indexing is incremental, so adding a single paper does not trigger a global rebuild.
• Query plannerRetrieval is a hybrid plan: BM25 for literal recall, dense retrieval for semantic recall, graph walks for neighbourhood expansion, and a learned reranker for the final ordering.
• CachingCommon neighbourhoods are cached per project; rarely traversed regions are materialised on demand. Project updates invalidate the relevant cache entries so new work can be picked up promptly.
• ObservabilityEvery graph traversal writes a provenance row you can inspect. Ask "why this paper?" and Academe can show the edges it used.

• Not a fixed ontology for all researchAcademe does not attempt to encode "all of science" as a fixed ontology. Concepts are soft labels, claims are extracted, and edges carry confidences. The graph is calibrated, not declarative.
• Not a replacement for readingThe graph is a map. It tells you where to look. Reading, interpreting, and judging papers is still your job. Academe helps keep relevant work visible.
• Not opinion-freeOur ranking and claim-extraction models have been trained and tuned by us; they make choices. We publish the benchmark results and edge-extraction accuracy so those choices are visible.
• Not the public internetNews articles, blog posts, social media, and unvetted grey sources are not in the graph unless a scholarly work explicitly references them. The index stays scholarly on purpose.