Skip to main content

​
What it does

Answers “what happened when?” questions over a dataset built by the ordinary cognify. Default cognify writes every date it extracts as a shared Timestamp node, linked to the chunks and entities that name it; the TEMPORAL search type runs hybrid retrieval and reranks the candidates by how well their Timestamps fit the time window the question asks about. This matches Python’s behaviour since cognee 1.6.3 (see Temporal Knowledge Graph).

​
There is no temporal cognify

Rust used to have a separate event-graph pipeline (--temporal-cognify, CognifyConfig::with_temporal_cognify, the bindings’ temporalCognify option). Python deprecated temporal_cognify and ignores it (cognee 1.6.2 was the last release in which temporal_cognify=True built the event graph). Rust removed it outright:
  • the CLI rejects --temporal-cognify as an unknown argument;
  • the bindings warn about temporalCognify as an unknown option and ignore it, as Python ignores the keyword (with a warning since cognee 1.6.5);
  • CognifyConfig has no temporal_cognify field, and data_per_batch, which only sized the temporal batches, is gone too.
Just run cognify. A dataset an older build cognified with the event pipeline holds Event nodes that nothing reads any more, so TEMPORAL answers on it with plain hybrid results. To get the rerank, rebuild it: forget the dataset’s memory and cognify again.

​
Querying

This uses SearchType::Temporal, served by TemporalHybridRetriever (crates/search/src/retrievers/temporal_hybrid/): hybrid retrieval reranked by the Timestamp nodes inside the question’s time window (see Operations). When the question names no time, or nothing in the graph is dated inside its window, the answer is the plain hybrid one. The recall router picks the temporal strategy when a time preposition comes right before an absolute date (in 2019, before July 1969) (crates/search/src/query_router.rs).

​
Pointers