For an end-to-end guide follow this link: https://www.timescale.com/blog/building-a-reverse-video-search-system-with-mixpeek-postgresql/

​
Video

We’ll be using vuse-generic-v1 to build a collection of 1 second interval video chunks into a 768 dimension embedding collection.

You’ll need to create a TimescaleDB database with the vector extension installed. We’ll create a hypertable called video_embeddings with a vector column of 768 dimensions and a time column for efficient time-series operations.

​
Ingest

from mixpeek import Mixpeek
import psycopg2
from psycopg2.extras import execute_values

# Initialize the Mixpeek client with your API key
mixpeek = Mixpeek("YOUR_API_KEY")

# Connect to TimescaleDB
conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

# Create table and hypertable if not exists
cur.execute("""
CREATE TABLE IF NOT EXISTS video_embeddings (
    id SERIAL PRIMARY KEY,
    embedding vector(768),
    start_time TIMESTAMPTZ,
    end_time TIMESTAMPTZ
);
""")
cur.execute("SELECT create_hypertable('video_embeddings', 'start_time', if_not_exists => TRUE);")

# Process video chunks
processed_chunks = mixpeek.tools.video.process(
    video_source="https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/starter/Jurassic+Park+(2).mp4",
    chunk_interval=1, # 1 second intervals
    resolution=[720, 1280]
)

embeddings_data = []
for chunk in processed_chunks:
    print(f"Processing video chunk: {chunk['start_time']}")

    # embed each chunk
    embed_response = mixpeek.embed.video(
        model_id="vuse-generic-v1",
        input=chunk['base64_chunk'],
        input_type="base64"
    )

    embeddings_data.append((
        embed_response['embedding'],
        chunk["start_time"],
        chunk["end_time"]
    ))

# Insert embeddings into TimescaleDB
execute_values(cur, """
    INSERT INTO video_embeddings (embedding, start_time, end_time)
    VALUES %s
""", embeddings_data)

conn.commit()
cur.close()
conn.close()

​
Text Query

query = "two boys inside a car"

embed_response = mixpeek.embed.video(
    model_id="vuse-generic-v1",
    input=query,
    input_type="text"
)

conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

cur.execute("""
    SELECT id, start_time, end_time, l2_distance(embedding, %s) AS distance
    FROM video_embeddings
    ORDER BY distance
    LIMIT 10
""", (embed_response['embedding'],))

results = cur.fetchall()
for result in results:
    print(result)

cur.close()
conn.close()

​
Video Query

# we'll use a cartoon version of jurassic park
file_url = "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/rabbit-jurassic.mp4"

embed_response = mixpeek.embed.video(
    model_id="vuse-generic-v1",
    input=file_url,
    input_type="url"
)

conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

cur.execute("""
    SELECT id, start_time, end_time, l2_distance(embedding, %s) AS distance
    FROM video_embeddings
    ORDER BY distance
    LIMIT 10
""", (embed_response['embedding'],))

results = cur.fetchall()
for result in results:
    print(result)

cur.close()
conn.close()

​
Image

We’ll be using openai-clip-vit-base-patch32 to build a collection of image embeddings with 512 dimensions.

You’ll need to create a TimescaleDB hypertable with a vector column of 512 dimensions for openai-clip-vit-base-patch32 model embeddings. We’re calling it image_embeddings.

​
Ingest

from mixpeek import Mixpeek
import psycopg2
from psycopg2.extras import execute_values

# Initialize the Mixpeek client with your API key
mixpeek = Mixpeek("YOUR_API_KEY")

# Connect to TimescaleDB
conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

# Create table and hypertable if not exists
cur.execute("""
CREATE TABLE IF NOT EXISTS image_embeddings (
    id SERIAL PRIMARY KEY,
    embedding vector(512),
    image_url TEXT,
    created_at TIMESTAMPTZ DEFAULT NOW()
);
""")
cur.execute("SELECT create_hypertable('image_embeddings', 'created_at', if_not_exists => TRUE);")

# List of image URLs to process
image_urls = [
    "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/image1.jpg",
    "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/image2.jpg",
    "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/image3.jpg",
    "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/image4.jpg"
]

embeddings_data = []
for url in image_urls:
    print(f"Processing image: {url}")

    # Embed each image
    embed_response = mixpeek.embed.image(
        model_id="openai-clip-vit-base-patch32",
        input=url,
        input_type="url"
    )

    embeddings_data.append((
        embed_response['embedding'],
        url
    ))

# Insert embeddings into TimescaleDB
execute_values(cur, """
    INSERT INTO image_embeddings (embedding, image_url)
    VALUES %s
""", embeddings_data)

conn.commit()
cur.close()
conn.close()

​
Text Query

query = "a cat sitting on a windowsill"

embed_response = mixpeek.embed.image(
    model_id="openai-clip-vit-base-patch32",
    input=query,
    input_type="text"
)

conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

cur.execute("""
    SELECT id, image_url, l2_distance(embedding, %s) AS distance
    FROM image_embeddings
    ORDER BY distance
    LIMIT 5
""", (embed_response['embedding'],))

results = cur.fetchall()
for result in results:
    print(result)

cur.close()
conn.close()

​
Image Query

# Use an image from the same bucket as a query
query_image = "https://mixpeek-public-demo.s3.us-east-2.amazonaws.com/timescaledb/query_image.jpg"

embed_response = mixpeek.embed.image(
    model_id="openai-clip-vit-base-patch32",
    input=query_image,
    input_type="url"
)

conn = psycopg2.connect("YOUR_TIMESCALEDB_CONNECTION_STRING")
cur = conn.cursor()

cur.execute("""
    SELECT id, image_url, l2_distance(embedding, %s) AS distance
    FROM image_embeddings
    ORDER BY distance
    LIMIT 5
""", (embed_response['embedding'],))

results = cur.fetchall()
for result in results:
    print(result)

cur.close()
conn.close()

Was this page helpful?

MongoDB