The building blocks, and how the pieces fit together.
Key concepts
A self-contained knowledge base: your sources plus everything curation derives from them. Contexts are independent — each has its own sources, manifest, index, and access. A query reads one or more contexts.
The plan for what to extract: a list of artifact types (each with a kind, scope, description, and — for structured types — columns) and the edge types that relate them. You shape it on Design; curation executes it.
The condensed knowledge curation produces. Every artifact has a kind:
summary A condensed overview — of the corpus or a single document.topic A theme or subject distilled across the sources.entity A named thing (account, person, product) with its relationships.event Something that happened, often time-stamped — good for timelines and counts.doc A per-document artifact, scoped to one source file.page A single source page — the finest grain, for exact quotes and figures.A query asks a question scoped to one or more curated contexts; Nexus plans its own retrieval and answers with citations. Multi-turn conversations are sessions — they carry their scope, system prompt, and history, and show the retrieval steps as Nexus works.
The programmatic counterpart to the console's query surface: query your contexts from your own code via the Query API. Same retrieval and citations — start or continue a session, stream the steps, and get a structured response. This is the integration path for applications — see Query API.
Under the hood, Nexus answers by writing Python against a retrieval SDK — reading artifacts off disk, running semantic/keyword search, walking the relationship graph, and querying structured tables — then synthesizing an answer from just the evidence it surfaced. This keeps raw tool output out of the model's context and lets one turn do a lot of work.