Building knowledge graphs with agents that verify every fact

Most existing knowledge graph construction pipelines process documents in batches without capturing semantic relationships across sections or verifying extracted information. RAGA combines an LLM agent with a structured toolkit for full knowledge graph operations (create, read, update, delete) and a Read-Search-Verify-Construct loop to build graphs while maintaining auditable provenance—every fact links back to source text. A hybrid retrieval mechanism merges symbolic graph queries with vector search. On QASPER (scientific QA), the approach outperforms zero-shot baselines and shows gains in both answer quality and evidence retrieval.