LAPIS

qdrant

3 min read

1. Setup

1.1 Run Qdrant

A simple way to start Qdrant locally is via Docker:

docker pull qdrant/qdrant  
docker run -p 6333:6333 -p 6334:6334 \
  -v "$(pwd)/qdrant_storage:/qdrant/storage:z" \
  qdrant/qdrant

This exposes the REST API on localhost:6333 and gRPC on localhost:6334, persisting data under ./qdrant_storage. Qdrant - Vector Database

Alternatively, you can spin up an in‑memory instance (for quick tests) directly from Python:

python


CopyEdit
from qdrant_client import QdrantClient  
client = QdrantClient(path=":memory:")  # no Docker needed

python-client.qdrant.tech

1.2 Install the Python Client

Install Qdrant’s official Python client via pip:

bash


CopyEdit
pip install qdrant-client

This provides CRUD and search methods for your collections. python-client.qdrant.tech

1.3 Install Jina Embeddings SDK

Use JinaAI’s Python SDK to generate text embeddings:

bash


CopyEdit
pip install jinaai

This gives you access to the jina-embeddings-v2-base-en (and other) models via a simple client API. PyPI

You can also install the full Jina framework if you plan to build pipelines beyond embeddings:

bash


CopyEdit
pip install jina

jina.ai

2. Ingesting Your Data

Assume you have a pandas DataFrame df with columns image_url, image_id, and caption.

2.1 Initialize Clients

python


CopyEdit
import pandas as pd
from qdrant_client import QdrantClient
from qdrant_client.models import PointStruct, Distance, VectorParams
from jinaai import EmbeddingClient

# Qdrant: connect to your local/memory instance
qdrant = QdrantClient(url="http://localhost:6333")  

# Jina: choose an embedding model
emb_client = EmbeddingClient(model="jina-embeddings-v2-base-en")

python-client.qdrant.techHugging Face

2.2 Create or Recreate a Collection

Define a collection named "images" with the appropriate vector size (e.g., 1024 for jie-embeddings-v2-base-en) and cosine distance:

python


CopyEdit
vector_dim = 1024
qdrant.recreate_collection(
    collection_name="images",
    vectors=VectorParams(size=vector_dim, distance=Distance.COSINE)
)

Qdrant - Vector Databasepython-client.qdrant.tech

2.3 Encode Captions and Prepare Points

Convert captions into vectors, then bundle each with its URL and ID as payload:

python


CopyEdit
# Load your DataFrame
# df = pd.read_csv("your_images.csv")

# Encode all captions at once
captions = df["caption"].tolist()
vectors = emb_client.encode(captions)  # returns List[List[float]]

# Build PointStruct list
points = [
    PointStruct(
        id=int(row.image_id),
        vector=vectors[i],
        payload={"image_url": row.image_url, "image_id": row.image_id}
    )
    for i, row in df.iterrows()
]

GitHubHugging Face

2.4 Upsert into Qdrant

Send your points in batches to Qdrant:

python


CopyEdit
qdrant.upsert(collection_name="images", points=points)

This will index your embeddings and store the associated payloads for filtering or retrieval. python-client.qdrant.tech

To find the top 10 images most similar to a new text query:

3.1 Embed the Query

python


CopyEdit
query_text = "sunset over mountains"
query_vector = emb_client.encode([query_text])[0]

Hugging Face

python


CopyEdit
results = qdrant.search(
    collection_name="images",
    query_vector=query_vector,
    limit=10
)

Each result includes the id, score, and your payload (with image_url and image_id). Qdrant - Vector Databasepython-client.qdrant.tech

Further Resources