What breaks in document AI

Indexing cost and re-indexing

8 of 8 quotes

• “the removed documents, together with their text units, entities, relationships and community reports, stay in the index and remain retrievable by queries.”

• “Input: multilingual (non-English) group-chat transcripts.”

• “update_communities leaves dangling entity_ids/relationship_ids after an incremental update”

• “We're running `Neo4JStorage` in production against AuraDB on a Graph-RAG retrieval workload over real-world enterprise KGs”

• “I run a production deployment of Graphiti (Neo4j on NAS, bge-m3 embeddings via Ollama, custom `memory-curator-daemon` doing incremental cycles + weekly Leiden community rebuild).”

• Hit this one too? We are building for it. Join the waitlist.

• “In production this translates to hours of unnecessary processing time and significant LLM API costs.”

• “the process took approximately 60 minutes to complete (from 12:29:34 to 13:29:46), averaging 36 seconds per record.”

• “We split into multiple subgraphs to stay within limits, but this breaks the "single graph" benefit, cross-graph relationships are lost”

• Hit this one too? We are building for it. Join the waitlist.