Unified Knowledge Graph RAG on AWS
An AWS-native framework that turns large multilingual document collections into knowledge graphs and answers multi-hop questions over them, unifying Microsoft's GraphRAG and LightRAG retrieval methods on Bedrock, Neptune and OpenSearch.
This framework reimplements two knowledge-graph retrieval methodologies, Microsoft's community-summarization GraphRAG and the dual-level keyword search of LightRAG, over one AWS-native stack, so a query can select either strategy while sharing a single ingestion, indexing, caching, multilingual and hybrid search infrastructure. GraphRAG search modes such as auto, drift, global, local and simple, and LightRAG modes such as mix, hybrid and naive, are chosen per query through a search-strategy parameter.
A DynamoDB document-status registry diffs an incoming corpus by content hash so incremental indexing re-indexes only new or changed documents and merges them into the live graph with idempotent upserts, while deletions remove only that document's exclusive graph artifacts. A prompt-tuning command profiles a corpus for domain, language, persona and entity types and emits domain-adapted prompts, a visualization command renders the graph from already-exported data without re-ingesting, and a graph-aware evaluator scores entity and relationship coverage. The project uses Bedrock, Neptune, OpenSearch, S3 and DynamoDB, follows a hexagonal, ports-and-adapters architecture so storage, retrieval, evaluation and rendering stay pluggable, and is Apache-2.0 licensed.