Graffiti tracks data lineage in LLM outputs

LLM systems can synthesize information from many inputs, producing summaries, extracted facts and structured records that may not appear verbatim in any original source. Daniel Chalef, founder of Zep AI and Graffiti, described this as a provenance problem because the path from source data to generated output can become unclear, complicating verification, debugging, legal compliance and user trust.

Simple source IDs are often inadequate for LLM pipelines because outputs may come from multiple evolving sources, while entities can merge and facts can change over time. Graffiti addresses this with an open-source temporal graph framework that models relationships between source data and derived artifacts. Facts can be connected to every episode that contributed to them, entity merges can unify lineage, and newer data can invalidate older facts while preserving the mutation history.

Graffiti also supports metadata projection, allowing source tags such as EHR or verified clinical source to be inherited by derived entities and facts. That enables filtering based on reliability, though multi-source facts still require policy decisions about whether any or all contributing sources must meet a verification threshold. The graph model also helps with deletion requests by removing facts derived solely from deleted data while preserving shared facts supported by remaining sources.