Semantic search

What is semantic search?

Semantic search is a technique that uses NLP to understand the intent and contextual meaning of a search query. This allows for results to be returned that are relevant to the user's intent, even if the query doesn't contain the exact keywords.

How does it work?

We use text embedding models like sentence-transformers, to transform your text into vectors that capture semantic information. These vectors represent the contextual meaning of sentences. When users make a search query, we encode their query into a vector and search for similar documents.

Indexing documents

We will rely on Elasticsearch to both transform certain documents fields into vectors and the user's query into a vector.

We will be indexing 4000 movies, vectorizing the plot field.

Load an embedding Model into Elasticsearch

In this example, we will use MiniLM-L6-v2 (opens in a new tab) from the sentence-transformers (opens in a new tab) library.

Follow this notebook (opens in a new tab) to do the steps below.

By the end of the notebook, you will have:

1. an embedding model running in Elasticsearch
2. an index that has a inference pipeline that uses the embedding model to transform the plot field into a vector
3. 4000 movies that been ingested into the index, with the plot field transformed into a vector

Create a Searchkit app

We will clone the Next.js example app and use this as a starting point.

curl https://codeload.github.com/searchkit/searchkit/tar.gz/main | \
tar -xz --strip=2 searchkit-main/examples/with-semantic-search-nextjs

Install dependencies

In the root of the project, install the dependencies:

yarn

Configure Searchkit

We will configure Searchkit to connect to Elasticsearch. This depends on how Elasticsearch is hosted. In this example, we are using Elastic Cloud.

 
const apiClient = API(
  {
    connection: {
      host: '<elasticsearch-host>',
      apiKey: '<api-key>'
    },
    search_settings: { ... }
  }
)
 

Update Display Fields

In this example, we are displaying the title and plot fields.

 
const apiClient = API(
  {
    connection: { ... },
    search_settings: {
      search_attributes: [],
      result_attributes: ['title', 'Plot'],
    }
  }
)
 

and also in the Hits component:

const HitView = (props: any) => {
  return (
    <div>
      <div className="hit__details">
        <h2>
          {props.hit.title}
        </h2>
        {props.hit.Plot}
      </div>
    </div>
  );
};

Configure Searchkit Query

We now need to update Searchkit's query to use perform a vector search instead of a regular search.

export async function POST(req: NextRequest, res: NextResponse) {
  const data = await req.json()
 
  const results = await apiClient.handleRequest(data, {
    // false to disable keyword search
    getQuery: () => false,
    getKnnQuery: (query) => {
      return {
        field: 'plot_embedding.predicted_value',
        k: 10,
        num_candidates: 50,
        query_vector_builder: {
          text_embedding: {
            model_id: 'sentence-transformers__all-minilm-l6-v2',
            model_text: query
          }
        }
      }
    }
  })
  return NextResponse.json(results)
}

Run the app

yarn dev

Hybrid Search

In this example, we are using a vector search to return the most relevant results. However, we can also combine this with a regular keyword search to return the most relevant results that match the query.

We start by defining search_attributes in the search_settings config.

 
const apiClient = API(
  {
    connection: { ... },
    search_settings: {
      search_attributes: ['title', 'Plot'],
      result_attributes: ['title', 'Plot'],
    }
  }
)
 

then we remove the getQuery function from the handleRequest call.

In the below example, when a customer searches, both the default getQuery and getKnnQuery will be called. The results from both queries will be combined and returned to the customer.

We will also send the rrf parameter to Elasticsearch to balance the different scores from the two queries.

export async function POST(req: NextRequest, res: NextResponse) {
  const data = await req.json()
 
  const results = await apiClient.handleRequest(data, {
    getKnnQuery: (query) => {
      return {
        field: 'plot_embedding.predicted_value',
        k: 10,
        num_candidates: 50,
        query_vector_builder: {
          text_embedding: {
            model_id: 'sentence-transformers__all-minilm-l6-v2',
            model_text: query
          }
        }
      }
    }
  })
  return NextResponse.json(results)
}