RAGA: LLM-Driven Autonomous Knowledge Graph Construction and Retrieval
RAGA (Reading And Graph-building Agent) introduces an innovative framework based on LLM for the autonomous creation and retrieval-augmented generation of knowledge graphs (KG). It overcomes the challenges posed by stateless batch processing methods, such as inadequate semantic relation capture across chunks, entity disambiguation issues, and lack of interpretability. RAGA offers a comprehensive toolset for executing all CRUD operations throughout the KG lifecycle and incorporates a cognitive constraint of Read-Search-Verify-Construct within a ReAct tool loop. Additionally, a KG-vector synchronization system facilitates hybrid symbolic-vector retrieval, while evidence-anchored verification connects knowledge entries to their source text, ensuring auditable provenance. Initial tests on a portion of the QASPER scientific QA dataset show enhanced retrieval fusion.
Key facts
- RAGA is an LLM-based autonomous KG construction and retrieval fusion framework.
- It addresses stateless batch processing pipeline limitations.
- Provides atomic toolset for full KG lifecycle CRUD operations.
- Embeds Read-Search-Verify-Construct cognitive constraint into ReAct tool loop.
- KG-vector synchronization mechanism enables hybrid symbolic-vector retrieval.
- Evidence-anchored verification links knowledge entries to source text.
- Preliminary experiments on QASPER scientific QA dataset subset.
- Paper available on arXiv with ID 2605.17072.
Entities
Institutions
- arXiv