GraphRAG

The core of GraphRAG is integrating knowledge graphs into the traditional RAG pipeline, organizing information in the form of triples (two entities and their relationship).

GraphRAG performs hierarchical clustering of knowledge graphs via community detection algorithms (such as the Leiden algorithm) to generate community summaries at different levels of granularity.

RAG（检索增强生成）是一种结合信息检索与生成式大模型的架构范式，模型生成答案前先根据用户查询从外部知识库检索最相关的文档片段，再将其作为上下文注入模型

The GraphRAG approach extracts entities and relationships from documents, stores them as nodes and edges in a graph database, and during retrieval first uses vector search to recall candidate nodes before performing multi-hop reasoning along the graph's edges.

Traditional RAG relies on chunk-based retrieval, splitting documents into segments and returning the most relevant passages through vector similarity matching, which cannot capture logical relationships and causal chains between entities.

Microsoft's 2024 research experiments demonstrated that GraphRAG significantly outperforms standard RAG on global questions requiring synthesis and reasoning across multiple documents, but with increased index construction costs and query latency.

Microsoft proposed GraphRAG in 2024, which represents an exploration in the direction of graph-based memory.

The GraphRAG architecture compensates for the limitations of pure vector retrieval by explicitly modeling character relationships and event causal chains through knowledge graphs.

Microsoft Research published the GraphRAG paper in 2024, demonstrating that GraphRAG significantly improves accuracy over traditional RAG when handling structured knowledge.

Skill Seeker的文档转Skills自动化方向与微软Research提出的GraphRAG思路相近

The solution employs agentic retrieval to give the retrieval process active reasoning capabilities, and uses the Cypher query language to perform structured queries on the graph.

GraphRAG Blueprint offers two retrieval modes—local mode and global mode—where local mode targets precise factual question answering and global mode is oriented toward cross-document synthesis and summarization.

GraphRAG organizes code knowledge by constructing a knowledge graph that uses functions, classes, and modules as nodes, and call relationships, inheritance relationships, and dependency relationships as edges.

The method introduces entity-driven GraphRAG retrieval, anchoring user questions on domain terms rather than relying solely on the closeness of semantic vectors.

GraphRAG is a variant of retrieval-augmented generation in which the retrieval stage first performs entity recognition and multi-hop traversal on a graph, then injects the matched subgraph as context into the generative model.

Modern AI search engines (such as Perplexity and ChatGPT Search) use vector space embeddings and graph-enhanced RAG retrieval, rather than crawling web pages in real time to count keyword density to determine visibility.

The structure of a knowledge triple is (subject entity)--[directed predicate]-->(object entity/value constraint), and it is the atomic data unit of GraphRAG.

Elastic believes graph databases (GraphRAG) have not been proven necessary in large-scale scenarios of billions of vectors, and currently prefers classification models and filtering mechanisms to solve relational reasoning problems.

FuseRAG achieves a 2.2 to 2.7 times TTFT speedup over the best GraphRAG baseline with no quality loss by separating reusable content from query-relevant structural interactions and injecting the latter into the decoder's self-attention layers.

KV-cache reuse presupposes a stable prefix, and GraphRAG flattens the graph structure into linear text, so the context prefix differs almost every query and the cache hit rate is extremely low.