MollyGraph

Local-first self-improving knowledge graph for AI agents with three-layer entity extraction, parallel graph and vector retrieval, and automated LoRA fine-tuning.

MollyGraph builds a knowledge graph that improves its own extraction model over time. It uses a three-layer NER pipeline (GLiNER2, spaCy enrichment, and GLiREL relation extraction) with speaker-anchored ingestion and per-source confidence thresholds. Queries run graph exact-match and vector similarity searches in parallel, merging and deduplicating results. The system includes an automated LoRA fine-tuning loop that only deploys new models when they beat the current one in A/B benchmarks, plus an LLM-powered audit chain for relationship verification. Backed by Neo4j and Jina embeddings, fully configured via environment variables.