Semantic Search With Dgraph and Modus

By leveraging embeddings and similarity search backed by a scalable vector index developers can enable semantic and similarity-based searches, improving the relevance of search results within their applications. This tutorial covers implementing semantic search using Modus, Dgraph, and Hypermode hosted AI models to add natural language or semantic search to your app using an example of ecommerce product data.

​

Semantic search overview

Semantic search focuses on understanding the meaning and context behind queries rather than just matching keywords, using embeddings to capture semantic relationships between concepts. Vector search serves as the technical implementation method, converting text into numerical vector embeddings and finding similar content through mathematical distance calculations in multidimensional space.

Vector search is a powerful technique that transforms data (like text, images, or audio) into numerical representations called embeddings. These embeddings capture the semantic meaning of the content in a multi-dimensional space and position similar items closer together. When performing a search, the query is also converted into an embedding, and the system finds items whose embeddings are closest to the query embedding.

This approach offers significant benefits over traditional keyword-based search, including improved relevance by capturing context and semantics, enhanced precision by understanding user intent, and the ability to handle complex queries with higher accuracy. Vector search is particularly effective for applications like semantic search, recommendation systems, and retrieval augmented generation (RAG), optimizing both efficiency and accuracy in finding and retrieving data based on meaningful similarity rather than exact matches. When combined with graph traversals, vector search can enable complex GraphRAG retrieval patterns.

​

The components

• Dgraph is a scalable graph database capable of near real-time graph traversals and vector search.
• Modus is the serverless API framework for building AI applications.
• Hypermode is the platform for deploying AI applications, including model hosting.

​

Prerequisites

This tutorial assumes you have:

1. Created a Dgraph instance, either hosted in the cloud or locally via Docker or installed via the Dgraph binary
2. Installed the Modus CLI and created a Modus app. See the Modus Quickstart to get started with Modus.

​

Natural language search with Dgraph and Modus

The steps to implement natural language or semantic search with Dgraph include defining the Dgraph connection in your Modus app manifest, selecting and configuring an embedding model, declaring a vector index in the Dgraph DQL schema, and using the similar_to DQL function to search for similar text in vector space.

Our example app uses ecommerce product data consisting of product details to enable semantic product search based on natural language terms.

​

Declare Dgraph connection and Hypermode embedding model

First, update the Modus app manifest file modus.json to define the connection to your Dgraph instance and the embedding model used to generate embeddings. Here we’re using the MiniLM model hosted by Hypermode and connecting to a locally running Dgraph instance.

{
  "$schema": "https://schema.hypermode.com/modus.json",
  "endpoints": {
    "default": {
      "type": "graphql",
      "path": "/graphql",
      "auth": "bearer-token"
    }
  },
  "models": {
    "minilm": {
      "sourceModel": "sentence-transformers/all-MiniLM-L6-v2",
      "connection": "hypermode",
      "provider": "hugging-face"
    }
  },
  "connections": {
    "dgraph-grpc": {
      "type": "dgraph",
      "grpcTarget": "localhost:9080"
    }
  }
}

​

Type definitions

Next, in our Modus app we define our data model using classes with decorators for automatic serialization/deserialization. The @json decorator enables JSON serialization, while @alias maps property names to Dgraph convention friendly formats.

We’ll be using ecommerce data so we’ll create simple types defining products and their categories.

@json
export class Product {
  @alias("Product.id")
  id!: string

  @alias("Product.title")
  title: string = ""

  @alias("Product.description")
  description: string = ""

  @alias("Product.category")
  @omitnull()
  category: Category | null = null
}

@json
export class Category {
  @alias("Category.name")
  name: string = ""
}

​

Embedding model integration

Next, we create an embedding function that uses a transformer model (MiniLM in this case) to convert product descriptions and search queries into vectors:

import { models } from "@hypermode/modus-sdk-as"
import { EmbeddingsModel } from "@hypermode/modus-sdk-as/models/experimental/embeddings"

const EMBEDDING_MODEL = "minilm"

export function embedText(content: string[]): f32[][] {
  const model = models.getModel<EmbeddingsModel>(EMBEDDING_MODEL)
  const input = model.createInput(content)
  const output = model.invoke(input)
  return output.predictions
}

​

Define Dgraph schema

We declare a schema to use Dgraph’s vector search capability and create an index on the Product.embedding property, even though Dgraph can function without a schema.

To define our Dgraph schema with vector indexing support we add the @index(hnsw) directive to the property storing the embedding value, in this case Product.embedding. We also define the other property types and node labels.

<Category.name>: string @index(hash) .
<Product.category>: uid @reverse .
<Product.description>: string .
<Product.id>: string @index(hash) .
<Product.embedding>: float32vector @index(hnsw) .

To apply this schema to our Dgraph instance we can make a POST request to the /alter endpoint of our Dgraph instance:

curl -X POST localhost:8080/alter --silent --data-binary '@dqlschema.txt'

or use the schema tab of the Ratel interface to apply the schema.

​

Define Modus mutation function

Now we’re ready to create a Modus function to create data in Dgraph. Here we create an upsert mutation that creates a product and related category in Dgraph, without creating duplicate nodes.

This function uses the embedding model we configured in previous steps to create an embedding of the product description and save to the Product.embedding property.

const DGRAPH_CONNECTION = "dgraph-grpc"

/**
 * Add or update a new product to the database
 */
export function upsertProduct(product: Product): Map<string, string> | null {
  let payload = buildProductMutationJson(DGRAPH_CONNECTION, product)

  const embedding = embedText([product.description])[0]
  payload = addEmbeddingToJson(payload, "Product.embedding", embedding)

  const mutation = new dgraph.Mutation(payload)
  const response = dgraph.executeMutations(DGRAPH_CONNECTION, mutation)

  return response.Uids
}

​

Dgraph similar_to query function

Next, we create a Modus function that uses Dgraph’s similar_to query function that leverages the vector index to find semantically similar products by computing an embedding of the search term and searching for nearby product descriptions in vector space.

/**
 * Search products by similarity to a given text
 */
export function searchProducts(search: string): Product[] {
  const embedding = embedText([search])[0]
  const topK = 3
  const body = `
    Product.id
    Product.description
    Product.title
    Product.category {
      Category.name
    }
  `
  return searchBySimilarity<Product>(
    DGRAPH_CONNECTION,
    embedding,
    "Product.embedding",
    body,
    topK,
  )
}

export function searchBySimilarity<T>(
  connection: string,
  embedding: f32[],
  predicate: string,
  body: string,
  topK: i32,
): T[] {
  const query = new dgraph.Query(`
    query search($vector: float32vector) {
        var(func: similar_to(${predicate},${topK},$vector))  {
            vemb as Product.embedding
            dist as math((vemb - $vector) dot (vemb - $vector))
            score as math(1 - (dist / 2.0))
        }

        list(func:uid(score),orderdesc:val(score))  @filter(gt(val(score),0.25)){
            ${body}
        }
    }`).withVariable("$vector", embedding)

  const response = dgraph.executeQuery(connection, query)
  console.log(response.Json)
  return JSON.parse<ListOf<T>>(response.Json).list
}

​

Query Modus endpoint

We can run our Modus app using the modus dev command which generates a GraphQL schema from the functions we’ve defined and start a local GraphQL endpoint for testing and development.

Navigate to http://localhost:8686/explorer in your browser and use the Modus API Explorer to first insert sample data into Dgraph using the upsert mutation function we defined previously and then search for similar products using vector search.

First, to create product and category nodes:

mutation ($product: ProductInput!) {
  upsertProduct(product: $product) {
    key
    value
  }
}

We’ll use the following values for the product variable creating three product nodes ad their associated category nodes in Dgraph:

And then to search for semantically similar products based on a search string we can execute the following query, using the value of our search string for the $search variable.

query ($search: String!) {
  searchProducts(search: $search) {
    id
    title
    description
    category {
      name
    }
  }
}

For example, if we search using the search term “rain”:

query {
  searchProducts(search: "rain") {
    id
    title
    description
    category {
      name
    }
  }
}

the product search results returns our solar powered umbrella.

{
  "data": {
    "searchProducts": [
      {
        "id": "P001",
        "title": "Solar-Powered Umbrella",
        "description": "A stylish umbrella with solar panels that charge your devices while you walk.",
        "category": {
          "name": "Outdoor Gear"
        }
      }
    ]
  }
}

Even though the description of the solar powered umbrella doesn’t include the word “rain” thanks to the meaning encoded into the embedding our semantic search process understands the association between rain and umbrella.

​

Next steps

Now that we’ve implemented semantic search using Dgraph, Modus, and Hypermode hosted models using the local development experience we’re ready to take the next step and deploy our project to Hypermode. See the Deploy Project section for a walk through of this process.

​

Resources

• You can find the full code for this example in the Modus Recipes GitHub repository: https://github.com/hypermodeinc/modus-recipes/tree/main/dgraph-101
• Watch a video overview of this tutorial in the Hypermode YouTube channel: https://www.youtube.com/watch?v=Z2fB-nBf4Wo