Vector Search

Vector search is a built-in capability in DataChain using cosine_distance(), euclidean_distance(), and l2_distance(). Create a vector (list of floats) and run searches against it.

import numpy as np
import datachain as dc
from PIL import Image
from sentence_transformers import SentenceTransformer

SRC = "gs://datachain-demo/dogs-and-cats/*jpg"
QUERY_LOCAL_IMG = "mycat.jpg"
model = SentenceTransformer("clip-ViT-B-32")


def clip_embedding(file: dc.ImageFile) -> list[float]:
    img = file.read().convert("RGB")
    emb = model.encode(img).astype(np.float32)
    emb = emb / (np.linalg.norm(emb) + 1e-12)
    return emb.tolist()


# 1) Build embeddings for all images in storage
ds = dc.read_storage(SRC, type="image", anon=True).map(emb=clip_embedding).persist()

# 2) Compute embedding for a local query image
query_img = Image.open(QUERY_LOCAL_IMG).convert("RGB")
query_emb = model.encode(query_img).astype(np.float32)
query_emb = query_emb / (np.linalg.norm(query_emb) + 1e-12)
query_emb = query_emb.tolist()

# 3) Similarity search (Top-10 closest images)
top10 = (
    ds.mutate(dist=dc.func.cosine_distance(dc.C("emb"), query_emb))
    .order_by("dist")
    .limit(10)
    .select("file.path", "dist")
)
top10.show()

The pattern: compute embeddings via map() (Python), then search via mutate + order_by + limit (Query Engine SQL). The similarity search runs at warehouse speed.

Vector distances work as analytical tools. Compare embeddings per record to reveal drift or disagreement:

import datachain as dc

chain.mutate(drift=dc.func.cosine_distance(dc.C("emb_a"), dc.C("emb_b"))).filter(
    dc.C("drift") > 0.3
).order_by("drift", descending=True)

This runs entirely in the Query Engine, with no Python and no deserialization.

Full Example: CLIP Embedding + Similarity Search

Model Drift Detection