How much of the knowledge base's graph a rebuild recomputes.
'full' — wipes the graph and re-extracts entities/relationships from every chunk with an LLM,
then rebuilds the concept layer. The expensive deep-clean: re-extraction dominates its cost.
'structural' — keeps the already-extracted entity graph (LLM extraction runs only for contexts
that were never graph-indexed, e.g. after an interrupted ingest) and recomputes the derived
structure on top of it: entity resolution, lexical mentions, topics and facets. Typically a small
fraction of a full rebuild's cost — the right choice for folding freshly ingested documents into
topics/facets.
How much of the knowledge base's graph a rebuild recomputes.
'full'— wipes the graph and re-extracts entities/relationships from every chunk with an LLM, then rebuilds the concept layer. The expensive deep-clean: re-extraction dominates its cost.'structural'— keeps the already-extracted entity graph (LLM extraction runs only for contexts that were never graph-indexed, e.g. after an interrupted ingest) and recomputes the derived structure on top of it: entity resolution, lexical mentions, topics and facets. Typically a small fraction of a full rebuild's cost — the right choice for folding freshly ingested documents into topics/facets.