Hi everybody. I’m Amir, founding father of Papr.
We constructed Papr Graph after seeing AI brokers fail in manufacturing. The mannequin wasn’t the issue — retrieval was once. Multi-hop questions, versioned insurance policies, relational information — flat vector seek breaks on it all.
Vector seek ranks by way of semantic closeness. However closeness ≠ correctness. A document announcing “aspirin reduces center assault possibility” and one announcing “aspirin reasons abdomen bleeding” rank just about equivalent — they are each about aspirin. For an agent creating a advice, that is the distinction between useful and damaging.
Papr Graph is a graph-native embedding that sits between your present embeddings and your agent. It encodes structured indicators — matter, time, intent, entities, the rest you outline — at once into your embedding, so score displays which means in context, no longer simply floor similarity. It is model-agnostic, works with no matter embeddings you might be already the use of.
We noticed Papr Graph reinforce present embeddings on MTEB (coding, scifact, finance duties) by way of 5-20%. On Stanford STaRK (MAG synthesized 10% dataset), Papr Graph leads retrieval fashions with 92% hit@5 accuracy.
Getting began is loose. Stay your present stack. Upload our plugin. Drop graph-native score into your present retrieval go with the flow with one API name.
