OPEN LINK ↗
// 65d agoBENCHMARK RESULT
ctxgraph tops Graphiti on relation extraction
ctxgraph is a local-first Rust context graph engine that uses GLiNER v2.1 via ONNX Runtime to extract entities and relations from engineering decision traces. In its 50-episode benchmark, the project reports better relation F1 and far lower latency than a Graphiti-style GPT-4o pipeline, while staying CPU-only and API-free.
// ANALYSIS
Narrow, schema-aware extraction looks better than prompt-driven graph construction when the ontology is fixed and label stability matters more than open-ended generation.
- –The reported 0.520 vs 0.315 F1 gap and 330ms vs 12.7s latency gap make this a practical systems win, not just a cost tweak.
- –Multiple GPT calls per episode compound latency and cost, so local inference pays off fast in agent memory loops.
- –Stable labels and direction validation matter more than raw recall; free-form edge names make graphs hard to query, dedupe, and trust.
- –The open benchmark corpus is valuable because it lets others test whether the gain comes from the model, the heuristics, or both.
// TAGS
ctxgraphbenchmarkinferencellmagentopen-sourceself-hosted
DISCOVERED
65d ago
2026-03-23
PUBLISHED
66d ago
2026-03-23
RELEVANCE
8/ 10
AUTHOR
synapse_sage