Tubegraph

Dgraph and its features

8 min read

From: aidotengineer

Dgraph is an open-source graph database first released in 2017, designed to optimize for large-scale graph data [00:04:23]. Its “D” stands for distributed, focusing on scaling in terms of data volume [00:04:37]. The original Dgraph team was inspired by Google’s Spanner graph [00:04:46].

The Knowledge Graph Mullet

Dgraph embodies the concept of the knowledge graph mullet, combining elements from both property graph and RDF triples worlds [00:01:01]. This hybrid approach offers a versatile way to work with knowledge graphs, leveraging the best of both paradigms [00:01:08].

Dgraph uses the property graph model for data modeling and querying, while employing RDF triples for data interchange, treating triples as the smallest unit of record [00:04:59].

Property Graph Model

The property graph model, which Dgraph uses for data modeling and querying, defines a knowledge graph as an instance of a property graph [00:02:58]. Key components include:

  • Nodes Nodes can have one or more labels that indicate their type or group them, similar to tables in relational databases [00:03:03].
  • Relationships Relationships have a single type and direction [00:03:24].
  • Properties Arbitrary key-value pair properties can be stored on both nodes and relationships [00:03:30].

The semantics, or how entities are connected, are encoded in the data model [00:03:36]. Instead of simply stating two nodes are related, Dgraph specifies the nature of their connection, such as a “talk has a certain topic” or “was presented at a conference” [00:03:40]. This approach focuses on connecting “things, not strings,” providing a canonical representation of entities [00:04:00].

RDF Triples

In the RDF world, the focus is on ontologies and triples (subject-predicate-object) [00:01:54]. RDF triples originated from the semantic web and linked data [00:02:05].

In Dgraph, a property graph is modeled as an RDF triple:

  • A unique ID for each node is required to model a property graph as an RDF triple [00:06:10].
  • The subject is always a node, specifically a unique ID pointing to the node [00:06:40].
  • The predicate can be a relationship or a property [00:06:49].
  • The object is either another node ID (if the predicate is a relationship) or the value of a property [00:06:53].

Optimization: Posting List

Dgraph employs an important optimization called a “posting list” [00:07:14]. This involves grouping by predicate and using a list of unique node IDs for all nodes connected by that predicate, allowing for efficient graph traversal [00:07:21].

DQL: Dgraph Query Language

DQL is the query language used with Dgraph, heavily inspired by GraphQL [00:07:38].

  • Traversal Structure Every DQL graph traversal begins with a well-defined starting point, often using an index to find root nodes [00:08:12].
  • Selection Set Queries use a nested selection set structure to specify properties to return and represent graph traversals [00:08:29].
  • JSON Output Data returned from a DQL query is in JSON format, matching the structure of the selection set [00:08:49].

Applications and Use Cases

News Knowledge Graph Example

To illustrate Dgraph’s capabilities, a news knowledge graph can be created [00:09:03].

  • Modeling Entities Entities mentioned in news articles, such as organizations, people, and topics, are modeled as nodes [00:09:16]. Authors, images, and unstructured data (like paragraphs) can also be represented as nodes [00:09:29].
  • Unstructured Data Handling Unstructured data is chunked (e.g., each paragraph as a chunk) and embedded [00:09:39]. The embedding of each chunk is calculated and stored as a node property, enabling vector search as an entry point into the graph [00:10:47].

Graph RAG Workflows

In a graph RAG approach, vector search (lexical graph) identifies relevant chunks, which serve as a starting point [00:11:20]. The system then traverses the graph (domain graph) from these chunks to related articles, topics, and organizations to gather relevant context for the model [00:11:36]. Other entry points include:

  • Geospatial Index Finding news articles within a certain region or near a specific location [00:11:55].
  • Image Similarity Search Using image embeddings as an entry point for searching [00:12:21].

This concept of different subgraph entry points is central to graph RAG [00:12:32].

Hands-on Example with Rattell

Rattell, a query workbench for Dgraph, allows users to execute DQL queries and visualize results [00:12:55]. Examples include:

  • Counting articles [00:13:10].
  • Searching for articles and traversing to connected topics [00:13:14].
  • Filtering articles by publication date and traversing to mentioned geographic areas [00:13:36].
  • Geographic distance search (e.g., finding news areas within 50 km of New York City) [00:13:57].
  • Vector search (e.g., finding articles close in vector space to a phrase like “money laundering”) [00:14:20].
  • Complex traversals combining vector search results with topic-based traversals to find related articles not caught by initial vector search [00:14:51].

Product Recommendations

Dgraph can be used for generating personalized product recommendations [00:22:01]. This is a typical graph database use case, where recommendations are generated by traversing the graph [00:22:08]. Approaches include:

  • Collaborative Filtering Finding similar users and identifying products they purchased that the target user has not [00:22:17].
  • Content-Based Recommendations Based on a user’s purchase history and product attributes [00:22:35].
  • Demographic Approaches Using demographic data to inform recommendations [00:22:43].

These approaches can be combined into a single database query [00:22:52].

Model Context Protocol (MCP) Server

Dgraph’s latest release includes a Model Context Protocol (MCP) server, which exposes tools to models, enabling them to interact with the database [00:15:37].

  • Dgraph MCP Server Each Dgraph instance serves an MCP server [00:16:50].
    • A read-only instance allows models to execute queries or inspect the schema [00:16:58].
    • A full endpoint also enables mutations (adding data) and altering the schema [00:17:06].
  • Use Cases
    • Agentic Coding Assistant Environments Tools like Windsurf or Cursor can leverage the schema or retrieved data from the MCP server to auto-generate code, such as CRUD endpoints or DQL queries [00:17:21].
    • Exploratory Data Analysis Tools like Claude desktop can generate DQL queries and fetch data to understand the graph’s contents [00:17:54].

Claude Desktop Integration

A demonstration shows integrating Dgraph’s MCP server with Claude desktop [00:18:15].

  1. Deploying Dgraph A Hypermode graph, powered by Dgraph, is deployed, including the MCP server endpoint [00:18:22].
  2. Configuring Claude The MCP configuration is copied into Claude desktop’s developer config [00:18:46].
  3. Schema and Data Generation Claude inspects the empty database, generates a graph schema (e.g., for e-commerce customer, product, and order data), and then generates a series of mutations to populate the graph with fictitious data [00:19:21].
  4. Model-Generated Queries The model generates database queries (DQL) and uses the MCP server tools to execute them against the database [00:20:00]. Claude can verify data creation and correct missing relationships through further mutations [00:20:16].
  5. Graph Visualization Claude can generate queries to fetch data and then produce JavaScript to render a graph visualization, aiding in understanding graph connections [00:21:07]. Different layout options (force-directed, hierarchical, radial) are available [00:21:41].
  6. Learning Tool This interaction serves as a valuable way to learn new developer tooling and query languages like DQL without needing to write them manually initially [00:20:45].

Modus Agent Orchestration Framework

Dgraph integrates with the open-source Modus agent orchestration framework, created by Hypermode to enable building agentic apps [00:24:07]. Modus provides abstractions for working with models and data, along with a runtime for managing large numbers of stateful, long-running AI agents [00:24:46]. It leverages WebAssembly for multi-language SDKs (e.g., Go, AssemblyScript) and generates a unified GraphQL API [00:25:16].

Hypermode Agents

Hypermode agents are powered by Dgraph and the Modus framework [00:26:21]. These domain-specific agents can be created starting with a prompt and by exposing tools via MCP servers [00:26:31]. Agents can interact with various services (e.g., GitHub, Notion, company docs via Ref) through their respective MCP servers [00:27:50].

Agent Demonstration

An example demonstrates a social media intern agent [00:27:13]:

  • Task Analyze a GitHub repo (Hypernews, containing Dgraph tooling for news knowledge graphs) and generate social media posts [00:28:17].
  • Tool Calls The agent searches for the repo, fetches file contents using GitHub MCP server tools, and generates posts [00:28:44].
  • Iteration The agent can be prompted to update posts to include relevant code snippets by accessing source files via GitHub tools [00:29:16].
  • Saving Output The agent can save the generated content to a Notion workspace page using Notion MCP server tools [00:30:00].
  • Code Ejection Users can “eject to code” to view the underlying Modus code running the agent, allowing for further customization and enhancement [00:31:18].

Hypermode agents are in early access [00:32:02].

Conclusion

Dgraph provides a powerful foundation for building AI agents and graph RAG workflows by combining the strengths of property graph and RDF triples ecosystems in a knowledge graph mullet approach [00:32:17]. Its features, including DQL, distributed architecture, and the MCP server, facilitate effective interaction between AI models and knowledge graphs.

Graph View