ZEHN Core
An enterprise knowledge graph GraphRAG question-answering system built on FalkorDB. It adopts a citation-first strategy, requiring answers to be traced to sources or else abstaining from answering.
Timeline (last 90 days)
The ZEHN Core project is built on FalkorDB, took part in the FalkorDB Graph Hacks hackathon, and its code is open-sourced.
ZEHN Core's design principle is that every answer must cite its source sentence, otherwise it answers 'I don't know'.
Retrieval is performed through whitelisted named queries, up to three hops, with a traversal policy deciding which predicates may be traversed.
The citation verification mechanism compares the set of each evidence ID output by the model against the IDs actually returned by the tool, and if any ID is not in the returned set, the entire answer is deemed invalid.
Forced abstention means the system proactively outputs 'I don't know' when valid evidence is lacking, and this gating is decided by code rather than by the model.
The system compares each hop of each returned path against the source sentences, labeling them as one of three categories: 'source-supported', 'inferred assumption', or 'insufficient evidence'.
When asked about the Redwood API, the graph found twelve paths, of which eleven were labeled as inferred assumptions.
The team's empirical testing found that 0 of 186 facts came verbatim from the source text.
Manual auditing found that about one third of the sampled relationships were bound to the wrong entity target.
In ZEHN Core's graph, every edge is physically of type RELATES, with the semantic type stored in the r.rel_type property, so querying :DEPENDS_ON returns empty.
3 more timeline events
All Facts (13)
The ZEHN Core project is built on FalkorDB, took part in the FalkorDB Graph Hacks hackathon, and its code is open-sourced.
50%UnverifiedZEHN Core's design principle is that every answer must cite its source sentence, otherwise it answers 'I don't know'.
50%UnverifiedRetrieval is performed through whitelisted named queries, up to three hops, with a traversal policy deciding which predicates may be traversed.
50%UnverifiedThe citation verification mechanism compares the set of each evidence ID output by the model against the IDs actually returned by the tool, and if any ID is not in the returned set, the entire answer is deemed invalid.
50%UnverifiedForced abstention means the system proactively outputs 'I don't know' when valid evidence is lacking, and this gating is decided by code rather than by the model.
50%UnverifiedThe system compares each hop of each returned path against the source sentences, labeling them as one of three categories: 'source-supported', 'inferred assumption', or 'insufficient evidence'.
50%UnverifiedWhen asked about the Redwood API, the graph found twelve paths, of which eleven were labeled as inferred assumptions.
50%UnverifiedThe team's empirical testing found that 0 of 186 facts came verbatim from the source text.
50%UnverifiedManual auditing found that about one third of the sampled relationships were bound to the wrong entity target.
50%UnverifiedIn ZEHN Core's graph, every edge is physically of type RELATES, with the semantic type stored in the r.rel_type property, so querying :DEPENDS_ON returns empty.
50%UnverifiedThe team evaluated making CALLS edges traversable, but because they cannot be supported by evidence, they remain disabled in production.
50%UnverifiedZEHN Core's corpus is a 30-document subset covering four source types, with no ticket source and no staleness detection.
50%UnverifiedZEHN Core is delivered with a FastAPI backend plus a React frontend, includes seven screens, and is backed by 959 backend tests and 63 frontend tests.
50%