Most "Graph RAG" implementations are vector retrieval with extra steps. @datagraphs built something...
TL;DR · AI 摘要
Qdrant 指出多数“图RAG”实为套壳向量检索，而 @datagraphs 构建了真正融合图数据库与向量搜索的协同架构：通过 schema-first agent 调度并行图查与语义检索，实现可验证、低延迟、全栈可控的知识问答。
核心要点
- 图与向量非互斥，而是互补：图擅精确逻辑查询（否定/时间/计算），向量擅语义相似匹配
- 真实图RAG需schema感知agent调度、实时图嵌入同步、图/向量并行检索与结果融合
- Qdrant被选中主因是混合云部署能力、payload过滤DSL兼容性及Terraform原生支持
结构提纲
按章节快速跳转。
思维导图
用一张图看清主题之间的关系。
查看大纲文本（无障碍 / 无 JS 友好）
- 真正协同的图RAG架构
- 核心理念
- 图与向量互补而非竞争
- 检索层决定AI智能上限
- 关键技术组件
- Proprietary图数据库引擎
- Qdrant语义搜索
- Agentic路径调度层
- 工程关键实践
- Schema-first检索
- 流式图嵌入同步
- 图/向量并行+融合
- 源追溯与可验证性
金句 / Highlights
值得收藏与分享的关键句。
Vector excels at semantic similarity over unstructured content. Graphs excel at empirical queries: negation, date ranges, mathematical operations on connected data.
Schema-first retrieval where the agent pulls the full graph schema before deciding how to query.
Parallel execution across graph queries and vector retrieval, blended for the LLM.
Why Qdrant specifically: Hybrid Cloud kept the entire stack inside their own AWS environment.
The retrieval layer determines the ceiling of your AI's intelligence. Choose the right components. Compose them well.
@datagraphs built something different.
Their UK-based knowledge graph platform combines a proprietary graph database engine with Qdrant-powered semantic search, orchestrated by an agentic layer that picks https://t.co/eg5Zb2mDl8" / X
Most "Graph RAG" implementations are vector retrieval with extra steps.
built something different. Their UK-based knowledge graph platform combines a proprietary graph database engine with Qdrant-powered semantic search, orchestrated by an agentic layer that picks the right retrieval path for each prompt. The insight from CEO Paul Wilton: vector and graph aren't competitors. They're complementary. Vector excels at semantic similarity over unstructured content. Graphs excel at empirical queries: negation, date ranges, mathematical operations on connected data. Force one to do the other's job and the system breaks. What they built: - Real-time embedding from the graph into Qdrant collections via streaming and queuing - Schema-first retrieval where the agent pulls the full graph schema before deciding how to query - Parallel execution across graph queries and vector retrieval, blended for the LLM - Every answer cited back to source for verifiable provenance Why Qdrant specifically: - Hybrid Cloud kept the entire stack inside their own AWS environment - Payload filtering DSL closely matched their existing OpenSearch patterns, so common metadata structures worked across the graph, OpenSearch, and Qdrant without translation - Terraform support and infrastructure automation fit their existing CI/CD workflows 18 months in production. Zero significant issues. The retrieval layer determines the ceiling of your AI's intelligence. Choose the right components. Compose them well. Full case study: qdrant.tech/blog/case-stud