How to conduct vector similarity semantic search with Elasticsearch

In this article, we’ll learn how to do vector similarity search using elasticsearch with an example. Before jumping into the tutorial, let’s brush up on our knowledge and familiarise the basics of elasticsearch semantic search, vector similarity, similarity search, etc. You’re welcome to skip the intro and jump to topics that interest you from the index below.

1. What is Elasticsearch?
2. What is Semantic search?
3. What is vector similarity?
4. Conducting semantic search
5. Using embeddings for similarity search
6. Tutorial: Implementing a QA system
7. Setting up Elasticsearch
8. Choosing an embedding model
9. Setting up the environment
10. Building a search API

What is Elasticsearch?

Elasticsearch is an open-source, distributed search engine for search and analytics built on Apache Lucene. It enables users to store, search, and analyze large volumes of data quickly and in near real-time. Elasticsearch is an underlying technology that powers applications with complex search features and requirements. 

Elasticsearch, along with its ecosystem of components known as Elastic Stack, has found applications in various areas, including simple searches on websites or documents, collecting and analyzing log data, data analysis, and visualization.

Semantic search

An easy way to perform a similarity search would be to rank documents based on how many words they share with the query. But a document may be similar to the query even if they have very few words in common. This is why semantic search is important. 

The semantic search involves a type of search often seen in search engines. It retrieves content after comprehending the intent and meaning of the user’s search query. Semantic search is much more advanced than traditional text and keyword match searches. Traditional keyword search does not consider lexical variants or conceptual matches to the user’s search phrase. If the precise wording used in the query cannot be found in the overall content, incorrect results will be provided to the user. 

This type of search is based on two concepts:

• Search intent of the user: This refers to the intention behind the user’s search. It involves decoding the reason behind the question. This reasoning could be anything from wanting more knowledge to finding a particular purchase item. By understanding the intent behind the query, search engines can retrieve the most accurate results for the users.
• Relationship between the words in the search phrase: It is essential to decoding the meaning of all the words together in the search phrase instead of the individual words. This means understanding the relationship between those words, thus displaying results conceptually similar to the user’s query.

Use-cases of Semantic Similarity Search:

Question-answering system: Given a collection of frequently asked questions, the search can find questions with the same meaning as the user’s new query. It can then provide stored results of similar questions.

• Image search – In a dataset of captioned images, it can find images whose caption is similar to the user’s description.
• Document content search – Allows searching through several documents to find one that matches the user’s requirements.
• Article search – In a collection of articles, it can return articles with a title closely related to the user’s query.

What is vector similarity?

Vector similarity measures the similarity between two vectors of an inner product space. It is measured by the cosine of the angle between two vectors and determines whether two vectors are pointing in roughly the same direction. It is often used to measure document similarity in text analysis.

Conducting semantic search

Semantic search can be implemented through a variety of different approaches. NLP specialists have developed a unique technique known as text embeddings. Text embedding involves converting words and sentences into fixed-size dense numeric vectors. This means that any kind of unstructured text can be converted to vectors. These vectors enable the understanding of the contextual meaning of the text and can be used to find the similarity between the user query and the web pages. If the text embeddings to two texts are similar, the two texts are semantically similar. These vectors can be indexed in Elasticsearch to perform semantic similarity searches.

Using embeddings for similarity search

Text embeddings can be used to retrieve questions that are similar to a user’s query. This is done through the following process-

• Each question is run through a sentence embedding model during indexing to produce a numeric vector.
• A query is run through the same sentence embedding model to produce a vector when a user enters a query. We calculate the similarity between each question and the query vector to rank the responses. When comparing embedding vectors, it is common to use cosine similarity.

Recalling Cosine Similarity

Cosine similarity is a metric used to measure the similarity of documents, irrespective of their size. Mathematically, it measures the cosine of the angle between two vectors projected in a multi-dimensional space. The cosine similarity is advantageous because even if the two similar documents are far apart by the Euclidean distance (due to the size of the document), chances are they may still be oriented closer together. The smaller the angle, the higher the cosine similarity.

Tutorial: Implementing a QA system

Let’s implement a simple question-answering search engine using elasticsearch and a sentence embedding architecture.

Setting up Elasticsearch

• Download and extract the archive

wget https://artifacts.elastic.co/downloads/elasticsearch/elasticsearch-7.9.1-linux-x86_64.tar.gz

tar -xzf elasticsearch-7.9.1-linux-x86_64.tar.gz

• Launch elasticsearch server

cd elasticsearch-7.9.1
./bin/elasticsearch

All done! Now let’s see if the elasticsearch node is up and running by sending an HTTP request to port 9200 (default port of es node).

• Run the following command from your terminal

curl -X GET "localhost:9200/?pretty"

The above command returns something similar to this:

{
  "name" : "desktop-pc",
  "cluster_name" : "elasticsearch",
  "cluster_uuid" : "Qr8lzRyZQnuNlcmp4P_OPB",
  "version" : {
    "number" : "7.9.1",
    "build_flavor" : "default",
    "build_type" : "tar",
    "build_hash" : "f27399d",
    "build_date" : "2020-03-26T06:34:37.794943Z",
    "build_snapshot" : false,
    "lucene_version" : "8.6.2",
    "minimum_wire_compatibility_version" : "1.2.3",
    "minimum_index_compatibility_version" : "1.2.3"
  },
  "tagline" : "You Know, for Search"
}

Note: If you are using a non Linux distribution, download your OS compatible package from Elasticsearch Download Page

Choosing an embedding model

We must generate embeddings for our textual information to perform a semantic search. Since we are dealing with questions, we will use a sentence embedding model to generate the embeddings. Sentence embedding techniques represent whole sentences and their semantic information as vectors. This helps understand the text’s context, intent, and other nuances. Some of the state-of-the-art sentence embedding techniques are:

• Doc2Vec
• SentenceBERT
• InferSent
• Universal Sentence Encoder

We will be using Universal Sentence Encoder for generating sentence embeddings. The universal encoder supports several downstream tasks and thus can be adopted for multi-task learning, i.e., the generated sentence embeddings can be used for multiple tasks like sentiment analysis, text classification, sentence similarity, etc.

Architecture

Universal Sentence Encoder is based on two encoders, Transformer and Deep Averaging Network(DAN). Multi-task learning is possible because both models are general-purpose.

The input sentence is tokenized according to the PTB method (Penn Treebank) and passed through one of these models. 

• Transformer

Transformer architecture was developed by Google in 2017. It leverages self-attention with multi blocks to learn the context-aware word representation.

• Deep Averaging Network (DAN)

The Deep Averaging Network (DAN) is a very simple model in which the word embeddings of the input text are simply averaged and then fed to a feed-forward neural network.

The transformer architecture performs better but requires more resources to train. Although DAN doesn’t work as well as the transformer architecture. The advantage of DAN is a simple model that requires fewer training resources.

Setting up the environment

• Step 1 – Clone the project repo

git clone https://github.com/adarsh-ops/semantic-qa.git
cd semantic-qa

• Step 2 – Install dependencies

pip install requirements.txt

Indexing the dataset

The dataset used in this project is from COVID-Q, a dataset of 1,690 questions about COVID-19 from thirteen online sources. The dataset is annotated by classifying questions into 15 question categories and by grouping questions that ask the same thing into 207 question classes.

First, index our dataset in the elasticsearch node to create a search space. 

Dataset Overview [download]

We will only use the dataset’s Question ID, Question and Answer fields.

• Step 1 – Define the elastic search index mapping

{
    "mappings": {
        "properties": {
            "question": {
                "type": "text"
            },
            "answer": {
                "type": "text"
            },
            "question_vec": {
                "type": "dense_vector",
                "dims": 512
            },
            "q_id": {
                "type": "long"
            }
        }
    }
}

a. Mappings – Defines the structure of the index

b. Properties – Defines the field/type relations

c. Fields –

• 1. question – text field to hold questions in the dataset
  2. answer – text field to answer the respective question
  3. question_vec – 512-dimensional vector representation (embedding) of the question
  4. q_id – Id of type long to represent the question id

• Step 2 – Define the required configs in config.py

# Universal Sentence Encoder Tf Hub url
MODEL_URL = "https://tfhub.dev/google/universal-sentence-encoder/4"

# Elasticsearch ip and port
ELASTIC_IP = "localhost"
ELASTIC_PORT = 9200
# Min score for the match
SEARCH_THRESH = 1.2

• Step 3 – Index the question vectors and answers

Run the dump_qa.py file to index the dataset at data/COVID-QA.csv

python dump_qa.py

This creates an index named “covid-qa” in the elasticsearch node with the mapping defined in step 1.

Code Breakdown

• Load the universal-sentence-encoder model

model = hub.load(config.MODEL_URL)

• Connect to the elasticsearch node

connect_elastic(config.ELASTIC_IP, config.ELASTIC_PORT)

• Read and clean the dataset

df = pd.read_csv("data/COVID-QA.csv")
df.dropna(inplace=True, subset=["Answers", "Question"])

• Index each QA pair in the dataset along with the Question ID and generate embedding into covid-qa index

for _, row in tqdm(df.iterrows()):
    insert_qa({
        'question': row['Question'],
        'answer': row['Answers'],
        'question_vec': np.asarray(model([row['Question']])[0]).tolist(),
        'q_id': row['Question ID']
    })

Here, the model([row[‘Question’]]) generates a 512-dimensional embedding for the given question.

Building a search API

With the search space created, all that is left to do is define the semantic search criteria. For this, let’s build an API for returning ‘top_n’ results given a search query (question).

• Step 1 – Define the API server and load configs

from flask import Flask, request

app = Flask(__name__)
app.config.from_object('config')

• Step 2 – Load the universal-sentence-encoder and connect it to es node

model = hub.load(app.config['MODEL_URL'])
connect_elastic(app.config['ELASTIC_IP'], app.config['ELASTIC_PORT'])

• Step 3 – Define the semantic search criteria

s_body = {
        "query": {
            "script_score": {
                "query": {
                    "match_all": {}
                },
                "script": {
                    "source": "cosineSimilarity(params.query_vector, 'question_vec') + 1.0",
                    "params": {"query_vector": query_vec}
                }
            }
        }
    }

For estimating the nearest ‘n’ records, cosine similarity between the query vector and the indexed question vectors are calculated. Ideally, the cosine similarity range is [-1, 1]. To change the score into real positive values, adding ‘1’ to the score will update the range to [0, 2]. 

result = es_conn.search(index="covid-qa", body=s_body)

for hit in result["hits"]["hits"]:
    print("--\nscore: {} \n question: {} \n answer: {}\n--".format(hit["_score"], hit["_source"]['question'], hit["_source"]['answer']))

• Step 4 – Define the API to perform a semantic search

@app.route("/query", methods=["GET"])
def qa():
    if request.args.get("query"):
        query_vec = np.asarray(model([request.args.get("query")])[0]).tolist()
        records = semantic_search(query_vec, app.config['SEARCH_THRESH'])
    else:
        return {"error": "Couldn't process your request"}, 422
    return {"data": records}

For performing the semantic vector match, we need to represent the raw text query as embeddings, model([request.args.get(“query”)]) generates a 512-dimensional embedding for the input query. 

• Step 5 – Run the API server

app.run(host="0.0.0.0", port=5000)

The server will be up and running on port 5000 of your machine.

So far, we’ve discussed semantic similarity, its applications, implementation techniques and built a simple QA engine using elasticsearch and a universal sentence encoder.

Semantic matching is helpful in applications like paraphrase identification, question answering, natural language generation, and intelligent tutoring systems. In short, if we want our system to be robust regarding grammatical reasoning, we would need to add a semantic analyzing technique. In some use cases, a hybrid approach may perform better. The QA engine we built earlier can be made more efficient by introducing text pre-processing techniques like cleansing, stopping word removal, etc.