Answer a broad question via community reports (map-reduce)
GraphRAG's actual "Global Search": every community report at `level` (default: the coarsest level generated, for the fewest, broadest reports and the lowest cost) gets one model call asking it for the key points it offers toward answering `question`, each with a 0-100 relevance score (the "map" step, run in parallel, capped at `max_communities` reports — the most similar to the question by embedding, when there are more than that at the level). The point across every report above a relevance floor are then pooled and given to one more model call that synthesizes them into a final answer (the "reduce" step). Suited to thematic or corpus-wide questions ("what are the main categories here?") that per-document retrieval answers poorly; for a question about one specific item, `createChatCompletion`'s ordinary retrieval is cheaper and more precise. Requires at least one `generateGraphCommunities` run for this collection.
Full-key operation
GraphRAG's actual "Global Search": every community report at level (default: the coarsest level generated, for the fewest, broadest reports and the lowest cost) gets one model call asking it for the key points it offers toward answering question, each with a 0-100 relevance score (the "map" step, run in parallel, capped at max_communities reports — the most similar to the question by embedding, when there are more than that at the level). The point across every report above a relevance floor are then pooled and given to one more model call that synthesizes them into a final answer (the "reduce" step). Suited to thematic or corpus-wide questions ("what are the main categories here?") that per-document retrieval answers poorly; for a question about one specific item, createChatCompletion's ordinary retrieval is cheaper and more precise. Requires at least one generateGraphCommunities run for this collection.
Authorization
api_key_header Every request carries an organization API key. A key has one of two scopes. A full key reaches every operation in this document and belongs on your server. A search key is the one to ship in a browser widget or a storefront page: it reaches only the operations marked x-nova-key-scope: search — searching, federated search, recommendations, event ingestion, and the visitor-facing AI operations (semantic and image search, voice transcription and voice search, assistant chat, and the visitor's own conversation). It cannot write documents or read or change configuration; any other operation answers 403. AI operations are paid in credits from the organization's wallet whichever key calls them (402 when the balance runs out), and a search key needs X-NOVA-SESSION-ID on them: the session is what a conversation belongs to, and what the per-session allowance (429) is counted against. A key of either scope may additionally carry restrictions set in the console: an allowlist of collections (403 elsewhere), a forced filter_by per collection that is ANDed into every search (document-level access control), an allowlist of fields (include_fields is limited to it and a request naming any other field in query_by, facet_by, sort_by, filter_by or highlight_fields answers 400), an IP/CIDR allowlist and an Origin/Referer allowlist (403), and a per-key rate limit (429 with Retry-After).
In: header
Path Parameters
The name of the collection
Request Body
application/json
TypeScript Definitions
Use the request body type in TypeScript.
Response Body
application/json
application/json
application/json
application/json
application/json
application/json
curl -X POST "https://example.com/collections/string/graph/global-search" \ -H "Content-Type: application/json" \ -d '{ "question": "string" }'{ "answer": "string", "level": 0, "communities_considered": 0, "sources": [ { "community_id": "string", "title": "string", "relevance_score": 0 } ]}Group entities into community reports POST
Groups this collection's extracted entities into a hierarchy of communities (multi-level Louvain modularity optimization over the relation graph — level 0 finest, each higher level a coarser merge of the level below, up to `max_levels`), and for every group of at least `min_size` entities at every level, asks a model for a short title and summary covering what the group is about. Existing reports for this collection are replaced. Re-run after `generateGraphEntities` adds enough new entities to be worth regrouping.
Answer a question via a Global Search primer plus Local Search follow-ups (DRIFT) POST
GraphRAG's "DRIFT" mode: a primer step (`globalSearchGraph`'s map-reduce over `level`'s community reports, at most `max_communities` of them) proposes an intermediate answer plus specific follow-up questions the reports can't answer on their own. Each of the first `max_follow_ups` follow-ups is then answered by a Local Search step — the entities most relevant to that follow-up by embedding similarity, their one-hop graph neighbors, and any claims about them (see `generateGraphEntities`) — and a final model call synthesizes the primer answer with every follow-up's local answer into one response. This is single-level DRIFT: follow-ups are answered but do not themselves spawn further follow-ups. Requires at least one `generateGraphCommunities` run; a collection with no entities yet still runs the primer step but every follow-up comes back with an empty `answer`.