EVENTFORGE: OPTIMIZABLE EVENT SCHEMA INDUCTION WITH HYBRID RAG+KG STORAGE

Abstract

Scientific literature, incident reports and technical documents encode process knowledge that is largely opaque to keyword search and modern retrieval-augmented generation. Event schema induction makes this knowledge available as a queryable graph, but current large-language-model inducers are often organized as hand-crafted prompt cascades that are brittle across models and difficult to improve from data. This paper presents EventForge, an open-source system that reframes schema induction as typed DSPy programs and co-locates the induced knowledge graph with a hierarchical dense-retrieval index in a single PostgreSQL/pgvector substrate. The system reads PDFs, extracts event candidates with DSPy ChainOfThought modules, removes near duplicates with sentence-embedding similarity, stores events in a NetworkX graph, and persists graph and retrieval artifacts together for hybrid RAG+KG querying. On a 12-PDF corpus spanning medical imaging, remote sensing, LLM security, and curriculum learning, EventForge induces 393 events and 286 graph relationships in 497.8 seconds on gpt-4o-mini, with a 6.2% near-duplicate rejection rate. The study reports system behavior, runtime, cost, graph yield, duplicate control, and implementation limitations, while leaving gold-standard event and edge F1 evaluation for future work.