Amazon S3 Vectors

Amazon S3 Vectors is a durable vector storage solution that can greatly reduce the total cost of uploading, storing, and querying vectors. S3 Vectors is a cloud object store with native support to store large vector datasets and provide subsecond query performance. This makes it more affordable for businesses to store AI-ready data at massive scale. This hands-on example walkthrough demonstrates how to use Amazon S3 Vectors with Unstructured. In this walkthrough, you will:

Create an S3 vector bucket.
Create a vector index in the bucket.
Add the contents of one or more source JSON output files that have been generated by Unstructured to the vector index.
Query the vector index against the contents of the source JSON output files that were added.

Requirements

To use this example, you will need:

A set of one or more JSON output files that have been generated by Unstructured and stored somewhere on your local development machine. For maximum compatibility with this example, these files must contain vector embeddings that were generated by Amazon Bedrock, by using the Titan Text Embeddings V2 (amazon.titan-embed-text-v2:0) embedding model, with 1024 dimensions. To get these files, you will need:
- An Unstructured account, as follows:
  - If you do not already have an Unstructured account, go to https://unstructured.io/contact and fill out the online form to indicate your interest.
  - If you already have an Unstructured account, sign in by using the URL of the sign in page that Unstructured provided to you when your Unstructured account was created. After you sign in, the Unstructured user interface (UI) then appears, and you can start using it right away. If you do not have this URL, contact Unstructured Sales at sales@unstructured.io.
- A workflow that generates vector embeddings and adds them to the JSON output files. Learn how to create a custom workflow and add an Embedder node to that workflow.
  The destination connector for your worklow must generate JSON output files. These include destination connectors for file storage connectors such as Databricks Volumes, Google Cloud Storage, OneDrive, and S3. Destination connectors for databases such as Elasticsearch, Kafka, and MongoDB, and vector stores such as Astra DB, Pinecone, and Weaviate, do not generate JSON output files.
- After your workflow generates the JSON output files, you must copy them from your workflow’s destination location over to some location on your local development machine for access.
Python installed on your local development machine.
An AWS account. Create an AWS account.

Step 1: Create the S3 vector bucket

Sign in to the AWS Management Console.
Open the Amazon S3 Console.
On the sidebar, click Vector bucket.
Click Create vector bucket.
For Vector bucket name, enter some name for your bucket.
For Encryption, select an encryption method, or leave the default.
Click Create vector bucket.

Step 2: Add the vector index to the bucket

With the list of vector buckets showing from the previous step, click the name of the bucket that you just created.
Click Create vector index.
For Vector index name, enter some name for your index.
For Dimension enter the number of dimensions that Unstructured generated for your vector embeddings. For example, for the Titan Text Embeddings V2 (amazon.titan-embed-text-v2:0) embedding model, enter 1024. If you are not sure how many dimensions to enter, see your workflow’s Embedder node settings.
Select the appropriate Distance metric for your embedding model. For example, for the Titan Text Embeddings V2 (amazon.titan-embed-text-v2:0) embedding model, select Cosine. If you are not sure which distance metric to use, see your embedding model’s documentation.
Expand Additional settings.
Within Metadata configuration, under Non-filterable metadata, click Add key.
For Key, enter text. This allows you to query the vector index by the text field within each object that will be coming over into the index from the JSON output files.
Click Create vector index.
After the vector index is created, copy the value of the index’s Amazon Resource Name (ARN), as you will need it in later steps. This ARN takes the format arn:aws:s3vectors:<region-id>:<account>:bucket/<bucket-name>/index/<index-name>.

Step 3: Add the source JSON output files’ contents to the vector index

In your local Python virtual environment, install the boto3 and uuid libraries.
Set up Boto3 credentials for your AWS account. The following steps assume you have set up your Boto3 credentials from outside of the following code, such as setting environment variables or configuring a shared credentials file, One approach to getting and setting up Boto3 credentials is to create an AWS access key and secret access key and then use the AWS Command Line Interface (AWS CLI) to set up your credentials on your local development machine.

Add the following code to a Python script file in your virtual environment, replacing the following placeholders:

Replace <source-json-file-path> with the path to the directory that contains your JSON output files.
Replace <index-arn> with the ARN of the vector index that you created previously in Step 2.
Replace <index-region-short-id> with the short ID of the region where your vector index is located, for example us-east-1.

import boto3
import os
import json
import uuid

source_json_file_path = '<source-json-file-path>'
index_arn             = '<index-arn>'
index_region_short_id = '<index-region-short-id>'
s3vectors             = boto3.client('s3vectors', region_name=index_region_short_id)
num_vectors           = 0
verbose               = True # Set to False to only print final results.

# For each JSON file in the source directory...
for filename in os.listdir(source_json_file_path):
    if filename.endswith('.json'):
        # ...read the JSON file, and then...
        with open(source_json_file_path + filename, 'r') as f:
            file = json.load(f)
            # ...for each object in the Unstructured-formatted JSON array...
            for object in file:
                # ...add the object's text and vectors to the S3 vector index.
                # Use the following format:

                # {
                #     "key": "<random-uuid>",
                #     "data": {
                #       "float32": <embedding-array>
                #     }, 
                #     "metadata": {
                #         "text": "<text-field>"
                #     }
                # }
                json_object = {}
                json_object['key'] = str(uuid.uuid4())
                json_object['data'] = {}
                json_object['metadata'] = {}

                # If the object has no text, do not add it to the 
                # vector index, and move on to the next object.
                if 'text' in object:
                    json_object['metadata']['text'] = object['text']
                else:
                    if verbose:
                        print(f"Skipping object with source element ID {object['id']} as it has no text")
                    continue

                # If the object has no embeddings, do not add it to the 
                # vector index either, and move on to the next object.
                if 'embeddings' in object:
                    json_object['data']['float32'] = object['embeddings']
                else:
                    if verbose:
                        print(f"Skipping object with source element ID {object['element_id']} as it has no embeddings")
                    continue

                # Add the object to the vector index.
                s3vectors.put_vectors(
                    indexArn=index_arn,
                    vectors=[json_object]
                )

                if verbose:
                    print(f"Added a vector entry and assigned it the internal ID {json_object['key']}. " + 
                          f"First 20 characters of the text: {object['text'][:20]}")
                
                num_vectors += 1

print(f"Added {num_vectors} vector entries to the vector index.")

Run the script to add the JSON output files’ contents to the vector index. Each object in each JSON output file is added as a vector entry in the vector index.

Step 4: Query the vector index

In your local Python virtual environment, install the numpy library.

Add the following code to another Python script file in your virtual environment, replacing the following placeholders:

Replace <index-arn> with the ARN of the vector index that you created previously in Step 2.
Replace <index-region-short-id> with the short ID of the region where your vector index is located, for example us-east-1.
Replace <sentence-to-embed> with the search text that you want to embed for the query.

import boto3
import json
import numpy as np

index_arn             = '<index-arn>'
index_region_short_id = '<index-region-short-id>'
client                = boto3.client('s3vectors', region_name=index_region_short_id)

# The sentence to embed.
sentence = '<sentence-to-embed>'

# Generate embeddings for the sentence to embed.
model_id    = 'amazon.titan-embed-text-v2:0'
bedrock     = boto3.client('bedrock-runtime', region_name=index_region_short_id)
body        = {'inputText': f"{sentence}"}
json_string = json.dumps(body)
json_bytes  = json_string.encode('utf-8')

response = bedrock.invoke_model(
    modelId=model_id,
    body=json_bytes,
    contentType='application/json',
    accept='application/json'
)

# Get the embeddings for the sentence and prepare them for the query.
response_body = json.loads(response['body'].read().decode())
embedding     = response_body['embedding']
embedding     = np.array(embedding, dtype=np.float32).tolist()

# Run the query.
response = client.query_vectors(
    indexArn=index_arn,
    topK=5,
    queryVector={'float32': embedding}
)

print(f"Original search query: {sentence}")
print(f"\nTop 5 results by similarity search...")
print('-----')

for vector in response['vectors']:
    response = client.get_vectors(
        indexArn=index_arn,
        keys=[vector['key']],
        returnData=True,
        returnMetadata=True
    )

    for vector in response['vectors']:
        print(vector['metadata']['text'])
        print('-----')

Run the script to query the vector index and see the query results.

Appendix: Additional operations

Use the following code examples to perform additional vector index and vector bucket operations.

List all entries in a vector index

Replace the following placeholders:

Replace <index-arn> with the ARN of the vector index that you created earlier in Step 2.
Replace <index-region-short-id> with the short ID of the region where your vector index is located, for example us-east-1.

import boto3

index_arn             = '<index-arn>'
index_region_short_id = '<index-region-short-id>'
client                = boto3.client('s3vectors', region_name=index_region_short_id)
num_vectors           = 0
next_token            = None
verbose               = True # Set to False to only print final results.

# List all of the vectors in the S3 vector index.
# Vectors are fetched by "page", so loop through the index in pages.
while True:
    kwargs = {
        'indexArn': index_arn,
        'returnData': True,
        'returnMetadata': True
    }

    if next_token:
        kwargs['nextToken'] = next_token

    response = client.list_vectors(**kwargs)

    for vector in response['vectors']:
        if verbose:
            print(f"Found vector entry with internal ID {vector['key']}. " +
                  f"First 20 characters of the text: {vector['metadata']['text'][:20]}")
        num_vectors += 1
    
    if 'nextToken' in response:
        next_token = response['nextToken']
    else:
        break

print(f"Total number of vector entries found: {num_vectors}")

Delete all entries from a vector index

This operation will permanently delete all vector entries in the vector index. This operation cannot be undone.

Replace the following placeholders:

Replace <index-arn> with the ARN of the vector index that you created earlier in Step 2.
Replace <index-region-short-id> with the short ID of the region where your vector index is located, for example us-east-1.

import boto3

index_arn             = '<index-arn>'
index_region_short_id = '<index-region-short-id>'
client                = boto3.client('s3vectors', region_name=index_region_short_id)
num_vectors           = 0
next_token            = None
verbose               = True # Set to False to only print final results.

# Delete all of the vectors in the S3 vector index.
# Vectors are deleted by "page", so loop through the index in pages.
while True:
    kwargs = {
        'indexArn': index_arn,
        'returnData': True,
        'returnMetadata': True
    }

    if next_token:
        kwargs['nextToken'] = next_token

    response = client.list_vectors(**kwargs)

    for vector in response['vectors']:
        # Delete each vector by its key.
        for vector in response['vectors']:
            if verbose:
                print(f"Deleting vector entry with internal ID {vector['key']}. " + 
                    f"First 20 characters of the text: {vector['metadata']['text'][:20]}")
                
            client.delete_vectors(
                indexArn=index_arn,
                keys=[vector['key']]
            )

            num_vectors += 1

    if 'nextToken' in response:
        next_token = response['nextToken']
    else:
        break

print(f"Deleted {num_vectors} vector entries from the vector index.")

Delete a vector index

This operation will permanently delete a vector index. This operation cannot be undone.

Replace the following placeholders:

Replace <index-arn> with the ARN of the vector index that you created earlier in Step 2.
Replace <index-region-short-id> with the short ID of the region where your vector index is located, for example us-east-1.

import boto3

index_arn             = '<index-arn>'
index_region_short_id = '<index-region-short-id>'
client                = boto3.client('s3vectors', region_name=index_region_short_id)

client.delete_index(
    indexArn=index_arn
)

Delete a vector bucket

This operation will permanently delete a vector bucket. This operation cannot be undone.

Replace the following placeholders:

Replace <bucket-arn> with the ARN of the vector bucket that you created earlier in Step 1. To get the ARN, do the following:
1. In the Amazon S3 console, on the sidebar, click Vector buckets.
2. Next to the name of the vector bucket that you want to delete, click the copy button next to the bucket’s Amazon Resource Name (ARN).
Replace <index-region-short-id> with the short ID of the region where your vector bucket is located, for example us-east-1.

import boto3

bucket_arn            = '<bucket-arn>'
index_region_short_id = '<index-region-short-id>'
client                = boto3.client('s3vectors', region_name=index_region_short_id)

client.delete_vector_bucket(
    vectorBucketArn=bucket_arn
)

Notebooks

Code samples

Tool demos

Requirements

Step 1: Create the S3 vector bucket

Step 2: Add the vector index to the bucket

Step 3: Add the source JSON output files’ contents to the vector index

Step 4: Query the vector index

Appendix: Additional operations

List all entries in a vector index

Delete all entries from a vector index

Delete a vector index

Delete a vector bucket

Notebooks

Code samples

Tool demos

​Requirements

​Step 1: Create the S3 vector bucket

​Step 2: Add the vector index to the bucket

​Step 3: Add the source JSON output files’ contents to the vector index

​Step 4: Query the vector index

​Appendix: Additional operations

​List all entries in a vector index

​Delete all entries from a vector index

​Delete a vector index

​Delete a vector bucket

Requirements

Step 1: Create the S3 vector bucket

Step 2: Add the vector index to the bucket

Step 3: Add the source JSON output files’ contents to the vector index

Step 4: Query the vector index

Appendix: Additional operations

List all entries in a vector index

Delete all entries from a vector index

Delete a vector index

Delete a vector bucket