Pgvector Vector Search Integration

Pgvector is a vector search extension to PostgreSQL, one of the most popular open source databases, and we’ve made it easy to use Pgvector on your computer vision data directly from FiftyOne!

Follow these simple instructions to get started using Pgvector + FiftyOne.

FiftyOne provides an API to create Pgvector indexes, upload vectors, and run similarity queries, both programmatically in Python and via point-and-click in the App.

Note

Did you know? You can search by natural language using Pgvector similarity indexes!

Basic recipe

The basic workflow to use Pgvector to create a similarity index on your FiftyOne datasets and use this to query your data is as follows:

1. Connect to or start a PostgreSQL server with the pgvector extension

2. Load a dataset into FiftyOne

3. Compute embedding vectors for samples or patches in your dataset, or select a model to use to generate embeddings

4. Use the compute_similarity() method to generate a Pgvector similarity index for the samples or object patches in a dataset by setting the parameter backend="pgvector" and specifying a brain_key of your choice

5. Use this Pgvector similarity index to query your data with sort_by_similarity()

6. If desired, delete the index

The example below demonstrates this workflow.

Note

You must have access to a PostgreSQL instance with pgvector extension and install the psycopg2 Python module to run this example:

pip install psycopg2

You can store credentials for your Postgres instance as described in this section to avoid entering them manually each time you interact with your Pgvector index.

First let’s load a dataset into FiftyOne and compute embeddings for the samples:

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

# Step 1: Load your data into FiftyOne
dataset = foz.load_zoo_dataset("quickstart")

# Steps 2 and 3: Compute embeddings and create a similarity index
pgvector_index = fob.compute_similarity(
    dataset,
    brain_key="pgvector_index",
    backend="pgvector",
)

Once the similarity index has been generated, we can query our data in FiftyOne by specifying the brain_key:

# Step 4: Query your data
query = dataset.first().id  # query by sample ID
view = dataset.sort_by_similarity(
    query,
    brain_key="pgvector_index",
    k=10,  # limit to 10 most similar samples
)

# Step 5 (optional): Cleanup

# Delete the pgvector index
pgvector_index.cleanup()

# Delete run record from FiftyOne
dataset.delete_brain_run("pgvector_index")

Note

Skip to this section for a variety of common pgvector query patterns.

Setup

The easiest way to get started with pgvector is to install locally via Docker.

Installing the psycopg2 client

In order to use the pgvector backend, you must also install the psycopg2 Python module:

pip install psycopg2

Using the pgvector backend

By default, calling compute_similarity() or sort_by_similarity() will use an sklearn backend.

To use the pgvector backend, simply set the optional backend parameter of compute_similarity() to "pgvector":

import fiftyone.brain as fob

fob.compute_similarity(..., backend="pgvector", ...)

Alternatively, you can permanently configure FiftyOne to use the pgvector backend by setting the following environment variable:

export FIFTYONE_BRAIN_DEFAULT_SIMILARITY_BACKEND=pgvector

or by setting the default_similarity_backend parameter of your brain config located at ~/.fiftyone/brain_config.json:

{
    "default_similarity_backend": "pgvector"
}

Authentication

If you are using a custom pgvector server, you can provide your credentials in a variety of ways.

Environment variables (recommended)

The recommended way to configure your pgvector credentials is to store them in the environment variables shown below, which are automatically accessed by FiftyOne whenever a connection to pgvector is made.

export FIFTYONE_BRAIN_SIMILARITY_PGVECTOR_CONNECTION_STRING=postgresql://postgres:mysecretpassword@localhost:5432/postgres

This is only one example of variables that can be used to authenticate a pgvector client. Find more information here.

FiftyOne Brain config

You can also store your credentials in your brain config located at ~/.fiftyone/brain_config.json:

{
    "similarity_backends": {
        "pgvector": {
            "connection_string": "postgresql://postgres:mysecretpassword@localhost:5432/postgres"
        }
    }
}

Note that this file will not exist until you create it.

Keyword arguments

You can manually provide credentials as keyword arguments each time you call methods like compute_similarity() that require connections to pgvector:

import fiftyone.brain as fob

pgvector_index = fob.compute_similarity(
    ...
    backend="pgvector",
    brain_key="pgvector_index",
    connection_string="postgresql://postgres:mysecretpassword@localhost:5432/postgres",
)

Note that, when using this strategy, you must manually provide the credentials when loading an index later via load_brain_results():

pgvector_index = dataset.load_brain_results(
    "pgvector_index",
    connection_string="postgresql://postgres:mysecretpassword@localhost:5432/postgres",
)

pgvector config parameters

The pgvector backend supports a variety of query parameters that can be used to customize your similarity queries. These parameters include:

• index_name (None): the name of the pgvector vector search index to use or create. If not specified, a new unique name is generated automatically

• table_name (None): the name of the postgres table to use or create for storing vectors. If not specified, a new unique name is generated automatically

• metric (“cosine”): the distance/similarity metric to use when creating a new index. The supported values are ("cosine", "dotproduct", "euclidean", "l1", "jaccard", "hamming")

• work_mem (“64MB”): the base maximum amount of memory to be used by a query operation (such as a sort or hash table) before writing to temporary disk files

• hnsw_m (16): The max number of connections per layer in the HNSW index

• hnsw_ef_construction (64): the size of the dynamic candidate list for constructing the graph for the HNSW index

For detailed information on these parameters, see the pgvector index options documentation.

You can specify these parameters via any of the strategies described in the previous section. Here’s an example of a brain config that includes all of the available parameters:

{
    "similarity_backends": {
        "pgvector": {
            "index_name": "your-index",
            "table_name": "your-table",
            "metric": "cosine",
            "work_mem": "64MB",
            "hnsw_m": 16,
            "hnsw_ef_construction": 64
        }
    }
}

However, typically these parameters are directly passed to compute_similarity() to configure a specific new index:

pgvector_index = fob.compute_similarity(
    ...
    backend="pgvector",
    brain_key="pgvector_index",
    index_name="your-index",
    table_name="your-table",
    metric="cosine",
    work_mem="64MB",
    hnsw_m=16,
    hnsw_ef_construction=64,
)

Managing brain runs

FiftyOne provides a variety of methods that you can use to manage brain runs.

For example, you can call list_brain_runs() to see the available brain keys on a dataset:

import fiftyone.brain as fob

# List all brain runs
dataset.list_brain_runs()

# Only list similarity runs
dataset.list_brain_runs(type=fob.Similarity)

# Only list specific similarity runs
dataset.list_brain_runs(
    type=fob.Similarity,
    patches_field="ground_truth",
    supports_prompts=True,
)

Or, you can use get_brain_info() to retrieve information about the configuration of a brain run:

info = dataset.get_brain_info(brain_key)
print(info)

Use load_brain_results() to load the SimilarityIndex instance for a brain run.

You can use rename_brain_run() to rename the brain key associated with an existing similarity results run:

dataset.rename_brain_run(brain_key, new_brain_key)

Finally, you can use delete_brain_run() to delete the record of a similarity index computation from your FiftyOne dataset:

dataset.delete_brain_run(brain_key)

Note

Calling delete_brain_run() only deletes the record of the brain run from your FiftyOne dataset; it will not delete any associated pgvector index, which you can do as follows:

# Delete the pgvector index
pgvector_index = dataset.load_brain_results(brain_key)
pgvector_index.cleanup()

Examples

This section demonstrates how to perform some common vector search workflows on a FiftyOne dataset using the pgvector backend.

Note

All of the examples below assume you have configured your pgvector server as described in this section.

Create a similarity index

In order to create a new pgvector similarity index, you need to specify either the embeddings or model argument to compute_similarity(). Here’s a few possibilities:

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")
model_name = "clip-vit-base32-torch"
model = foz.load_zoo_model(model_name)
brain_key = "pgvector_index"

# Option 1: Compute embeddings on the fly from model name
fob.compute_similarity(
    dataset,
    model=model_name,
    backend="pgvector",
    brain_key=brain_key,
)

# Option 2: Compute embeddings on the fly from model instance
fob.compute_similarity(
    dataset,
    model=model,
    backend="pgvector",
    brain_key=brain_key,
)

# Option 3: Pass precomputed embeddings as a numpy array
embeddings = dataset.compute_embeddings(model)
fob.compute_similarity(
    dataset,
    embeddings=embeddings,
    backend="pgvector",
    brain_key=brain_key,
)

# Option 4: Pass precomputed embeddings by field name
dataset.compute_embeddings(model, embeddings_field="embeddings")
fob.compute_similarity(
    dataset,
    embeddings="embeddings",
    backend="pgvector",
    brain_key=brain_key,
)

Create a patch similarity index

You can also create a similarity index for object patches within your dataset by including the patches_field argument to compute_similarity():

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

fob.compute_similarity(
    dataset,
    patches_field="ground_truth",
    model="clip-vit-base32-torch",
    backend="pgvector",
    brain_key="pgvector_patches",
)

Connect to an existing index

If you have already created a pgvector index storing the embedding vectors for the samples or patches in your dataset, you can connect to it by passing the index_name to compute_similarity():

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

fob.compute_similarity(
    dataset,
    model="clip-vit-base32-torch",      # zoo model used (if applicable)
    embeddings=False,                   # don't compute embeddings
    index_name="your-index",            # the existing pgvector index
    brain_key="pgvector_index",
    backend="pgvector",
)

Add/remove embeddings from an index

You can use add_to_index() and remove_from_index() to add and remove embeddings from an existing pgvector index.

These methods can come in handy if you modify your FiftyOne dataset and need to update the pgvector index to reflect these changes:

import numpy as np

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

pgvector_index = fob.compute_similarity(
    dataset,
    model="clip-vit-base32-torch",
    brain_key="pgvector_index",
    backend="pgvector",
)
print(pgvector_index.total_index_size)  # 200

view = dataset.take(10)
ids = view.values("id")

# Delete 10 samples from a dataset
dataset.delete_samples(view)

# Delete the corresponding vectors from the index
pgvector_index.remove_from_index(sample_ids=ids)

# Add 20 samples to a dataset
samples = [fo.Sample(filepath="tmp%d.jpg" % i) for i in range(20)]
sample_ids = dataset.add_samples(samples)

# Add corresponding embeddings to the index
embeddings = np.random.rand(20, 512)
pgvector_index.add_to_index(embeddings, sample_ids)

print(pgvector_index.total_index_size)  # 210

Retrieve embeddings from an index

You can use get_embeddings() to retrieve embeddings from a pgvector index by ID:

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

pgvector_index = fob.compute_similarity(
    dataset,
    model="clip-vit-base32-torch",
    brain_key="pgvector_index",
    backend="pgvector",
)

# Retrieve embeddings for the entire dataset
ids = dataset.values("id")
embeddings, sample_ids, _ = pgvector_index.get_embeddings(sample_ids=ids)
print(embeddings.shape)  # (200, 512)
print(sample_ids.shape)  # (200,)

# Retrieve embeddings for a view
ids = dataset.take(10).values("id")
embeddings, sample_ids, _ = pgvector_index.get_embeddings(sample_ids=ids)
print(embeddings.shape)  # (10, 512)
print(sample_ids.shape)  # (10,)

Querying a pgvector index

You can query a pgvector index by appending a sort_by_similarity() stage to any dataset or view. The query can be any of the following:

• An ID (sample or patch)

• A query vector of same dimension as the index

• A list of IDs (samples or patches)

• A text prompt (if supported by the model)

import numpy as np

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

fob.compute_similarity(
    dataset,
    model="clip-vit-base32-torch",
    brain_key="pgvector_index",
    backend="pgvector",
)

# Query by vector
query = np.random.rand(512)  # matches the dimension of CLIP embeddings
view = dataset.sort_by_similarity(query, k=10, brain_key="pgvector_index")

# Query by sample ID
query = dataset.first().id
view = dataset.sort_by_similarity(query, k=10, brain_key="pgvector_index")

# Query by a list of IDs
query = [dataset.first().id, dataset.last().id]
view = dataset.sort_by_similarity(query, k=10, brain_key="pgvector_index")

# Query by text prompt
query = "a photo of a dog"
view = dataset.sort_by_similarity(query, k=10, brain_key="pgvector_index")

Note

Performing a similarity search on a DatasetView will only return results from the view; if the view contains samples that were not included in the index, they will never be included in the result.

This means that you can index an entire Dataset once and then perform searches on subsets of the dataset by constructing views that contain the images of interest.

Advanced usage

As previously mentioned, you can customize your pgvector indexes by providing optional parameters to compute_similarity().

Here’s an example of creating a similarity index backed by a customized pgvector index. Just for fun, we’ll specify a custom index name, use dot product similarity, and populate the index for only a subset of our dataset:

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

# Create a custom pgvector index
pgvector_index = fob.compute_similarity(
    dataset,
    model="clip-vit-base32-torch",
    embeddings=False,  # we'll add embeddings below
    metric="dotproduct",
    brain_key="pgvector_index",
    backend="pgvector",
    index_name="custom-quickstart-index",
)

# Add embeddings for a subset of the dataset
view = dataset.take(10)
embeddings, sample_ids, _ = pgvector_index.compute_embeddings(view)
pgvector_index.add_to_index(embeddings, sample_ids)