Query-time knowledge graphs without the upkeep headache

What it does SAG ingests raw text, decomposes it into “semantic atomic events,” and extracts weighted multi-dimensional entities such as Time, Location, Person, Topic, Action, and Tags. At search time, it dynamically constructs a relationship network across those events and ranks results through a three-stage pipeline: entity recall, multi-hop BFS expansion, and PageRank reranking. The system stores structured data in MySQL and vectors in Elasticsearch or another VecDB.

The interesting bit The project treats pre-built knowledge graphs as too expensive to maintain. It flips the GraphRAG model by keeping events and entities in SQL tables, then weaving the graph edges on demand during retrieval. A configurable “5W1H” entity schema with per-dimension weights lets you tune how the engine interprets and connects concepts.

Key highlights

• Query-time graph construction: no static knowledge graph to curate or keep updated
• Three-stage retrieval: Recall → Expand (configurable-depth BFS) → Rerank with PageRank
• Custom entity dimensions: extend the default 5W1H schema with domain-specific types and weights
• Hybrid storage layer: MySQL for events and entity relations, Elasticsearch/VecDB for vector retrieval
• Open core: the retrieval engine is Apache 2.0, though team features and managed hosting are reserved for a commercial tier at zleap.ai

Caveats

• Several practical features—team collaboration, auto-syncing data sources, and one-click report generation—are gated behind the commercial “complete version”; the open-source release is single-user and requires self-hosting.
• The repository metadata lists TypeScript, yet the README and SDK examples center on Python and FastAPI, which suggests the open-source engine is Python-based with a TypeScript frontend or the metadata is simply misleading.
• Accuracy and scalability claims in the README rely on self-graded star charts rather than independent benchmarks.

Verdict A solid candidate if you want GraphRAG-style relationship search without the operational burden of a persistent graph database. Look elsewhere if you need a managed, multi-user knowledge base out of the box.