← back
Citation Needed: Provenance for LLM-Built Knowledge Graphs — Daniel Chalef, Zep AI
Takeaway
Build provenance into the evolving knowledge graph so every derived fact remains traceable through merges, invalidation, and deletion.
Summary
- Zep's Graphiti framework models provenance as relationships between source episodes, extracted entities, and facts because LLM synthesis cannot be traced with a single static source pointer.
- Entity merges must preserve all source links, while contradictions invalidate facts with timestamps and record which episodes caused the change.
- Source metadata propagates to derived artifacts, but the application must decide whether any or all supporting sources need a particular trust classification.
- When a source is deleted, facts with other supporting episodes can survive while unsupported facts are removed; provenance also enables debugging and inspection of source credibility.
graphitiprovenanceknowledge-graphs
Original description
An agent hands a doctor a clean, confident fact: the patient has a penicillin allergy. But that fact was synthesized from three sources, an EHR record, a lab report, and something the patient typed into an intake chatbot, and by the time it reaches the doctor, which one it came from is gone. You cannot just stamp a source ID on it, because the LLM merged entities and later data invalidated earlier facts, so the store keeps shifting under your pointer. Daniel Chalef's argument is that provenance for a knowledge graph an LLM builds has to be a graph itself. In Graphiti, the open source framework behind Zep, sources become episodes and every derived fact links back to them, so tracing a fact to its origin is just a graph walk. Tag a source once and the tag follows every node and edge derived from it, which lets an agent keep only facts from verified clinical sources. Deletion is the same walk in reverse: a GDPR erasure removes a source, and a fact survives only if another source still supports it. Compliance gets an audit trail, and engineers get agents they can debug instead of black boxes. Speaker info: https://x.com/danielchalef / danielchalef https://github.com/getzep/graphiti Timestamps: 0:00 - Why LLM synthesis destroys the paper trail 1:10 - Graphiti, Zep, and the provenance problem 1:47 - The failure mode: a penicillin allergy from three sources 2:53 - Why a source ID does not survive an LLM pipeline 4:20 - Provenance as a graph: tracing a fact is a walk 5:09 - Keeping lineage correct through merges and invalidation 6:06 - Metadata projection: tag a source once 7:25 - Mixed trust parents: allergy flags versus consent 8:57 - Deletion: GDPR erasure through the same edges 10:26 - Benefits: compliance, veracity, and debuggability 11:31 - Q&A: cost, dedup, and why not just markdown