Skip to main content

Batch import

Batch imports are an efficient way to add multiple data objects and cross-references.

Additional information

To create a bulk import job, follow these steps:

  1. Initialize a batch object.
  2. Add items to the batch object.
  3. Ensure that the last batch is sent (flushed).

Basic import​

The following example adds objects to the MyCollection collection.

data_rows = [
    {"title": f"Object {i+1}"} for i in range(5)
]

collection = client.collections.get("MyCollection")

with collection.batch.dynamic() as batch:
    for data_row in data_rows:
        batch.add_object(
            properties=data_row,
        )
py docs  API docs

Use the gRPC API​

Added in v1.23.

The gRPC API is faster than the REST API. Use the gRPC API to improve import speeds.

The Python client uses gRPC by default. See the client page for additional batch import configuration options.


The legacy Python client does not support gRPC.

Specify an ID value​

Weaviate generates an UUID for each object. Object IDs must be unique. If you set object IDs, use one of these deterministic UUID methods to prevent duplicate IDs:

from weaviate.util import generate_uuid5  # Generate a deterministic ID

data_rows = [{"title": f"Object {i+1}"} for i in range(5)]

collection = client.collections.get("MyCollection")

with collection.batch.dynamic() as batch:
    for data_row in data_rows:
        obj_uuid = generate_uuid5(data_row)
        batch.add_object(
            properties=data_row,
            uuid=obj_uuid
        )
py docs  API docs

Specify a vector​

Use the vector property to specify a vector for each object.

data_rows = [{"title": f"Object {i+1}"} for i in range(5)]
vectors = [[0.1] * 1536 for i in range(5)]

collection = client.collections.get("MyCollection")

with collection.batch.dynamic() as batch:
    for i, data_row in enumerate(data_rows):
        batch.add_object(
            properties=data_row,
            vector=vectors[i]
        )
py docs  API docs

Specify named vectors​

Added in v1.24

When you create an object, you can specify named vectors (if configured in your collection).

data_rows = [{
    "title": f"Object {i+1}",
    "body": f"Body {i+1}"
} for i in range(5)]

title_vectors = [[0.12] * 1536 for _ in range(5)]
body_vectors = [[0.34] * 1536 for _ in range(5)]

collection = client.collections.get("MyCollection")

with collection.batch.dynamic() as batch:
    for i, data_row in enumerate(data_rows):
        batch.add_object(
            properties=data_row,
            vector={
                "title": title_vectors[i],
                "body": body_vectors[i],
            }
        )
py docs  API docs

Import with references​

You can batch create links from an object to another other object through cross-references.

collection = client.collections.get("Author")

with collection.batch.fixed_size(batch_size=100) as batch:
    batch.add_reference(
        from_property="writesFor",
        from_uuid=from_uuid,
        to=target_uuid,
    )
py docs  API docs

Python-specific considerations​

The Python clients have built-in batching methods to help you optimize import speed. For details, see the client documentation:

Async Python client and batching​

Currently, the async Python client does not support batching. To use batching, use the sync Python client.

Stream data from large files​

If your dataset is large, consider streaming the import to avoid out-of-memory issues.

To try the example code, download the sample data and create the sample input files.

Get the sample data
import requests

# Download the json file
response = requests.get(
    "https://raw.githubusercontent.com/weaviate-tutorials/intro-workshop/main/data/jeopardy_1k.json"
)

# Write the json file to disk
data = response.json()
with open('jeopardy_1k.json', 'w') as f:
    json.dump(data, f)

# # Uncomment this section to create a csv file
# import pandas as pd

# df = pd.read_json("jeopardy_1k.json")
# df.to_csv("jeopardy_1k.csv", index=False)
py docs  API docs
Stream JSON files example code
import ijson

# Settings for displaying the import progress
counter = 0
interval = 200  # print progress every this many records; should be bigger than the batch_size

print("JSON streaming, to avoid running out of memory on large files...")
with client.batch.fixed_size(batch_size=100) as batch:
    with open("jeopardy_1k.json", "rb") as f:
        objects = ijson.items(f, "item")
        for obj in objects:
            properties = {
                "question": obj["Question"],
                "answer": obj["Answer"],
            }
            batch.add_object(
                collection="JeopardyQuestion",
                properties=properties,
                # If you Bring Your Own Vectors, add the `vector` parameter here
                # vector=obj.vector["default"]
            )

            # Calculate and display progress
            counter += 1
            if counter % interval == 0:
                print(f"Imported {counter} articles...")


print(f"Finished importing {counter} articles.")
py docs  API docs
Stream CSV files example code
import pandas as pd

# Settings for displaying the import progress
counter = 0
interval = 200  # print progress every this many records; should be bigger than the batch_size

def add_object(obj) -> None:
    global counter
    properties = {
        "question": obj["Question"],
        "answer": obj["Answer"],
    }

    with client.batch.fixed_size(batch_size=100) as batch:
        batch.add_object(
            collection="JeopardyQuestion",
            properties=properties,
            # If you Bring Your Own Vectors, add the `vector` parameter here
            # vector=obj.vector["default"]
        )

        # Calculate and display progress
        counter += 1
        if counter % interval == 0:
            print(f"Imported {counter} articles...")


print("pandas dataframe iterator with lazy-loading, to not load all records in RAM at once...")
with client.batch.fixed_size(batch_size=200) as batch:
    with pd.read_csv(
        "jeopardy_1k.csv",
        usecols=["Question", "Answer", "Category"],
        chunksize=100,  # number of rows per chunk
    ) as csv_iterator:
        # Iterate through the dataframe chunks and add each CSV record to the batch
        for chunk in csv_iterator:
            for index, row in chunk.iterrows():
                properties = {
                    "question": row["Question"],
                    "answer": row["Answer"],
                }
                batch.add_object(
                    collection="JeopardyQuestion",
                    properties=properties,
                    # If you Bring Your Own Vectors, add the `vector` parameter here
                    # vector=obj.vector["default"]
                )

        # Calculate and display progress
        counter += 1
        if counter % interval == 0:
            print(f"Imported {counter} articles...")

print(f"Finished importing {counter} articles.")
py docs  API docs

Batch vectorization​

Added in v1.25.

Some model providers provide batch vectorization APIs, where each request can include multiple objects.

From Weaviate v1.25.0, a batch import automatically makes use of the model providers' batch vectorization APIs where available. This reduces the number of requests to the model provider, improving throughput.

Model provider configurations​

You can configure the batch vectorization settings for each model provider, such as the requests per minute or tokens per minute. The following examples sets rate limits for Cohere and OpenAI integrations, and provides API keys for both.

Note that each provider exposes different configuration options.

from weaviate.classes.config import Integrations

integrations = [
    # Each model provider may expose different parameters
    Integrations.cohere(
        api_key=cohere_key,
        requests_per_minute_embeddings=rpm_embeddings,
    ),
    Integrations.openai(
        api_key=openai_key,
        requests_per_minute_embeddings=rpm_embeddings,
        tokens_per_minute_embeddings=tpm_embeddings,   # e.g. OpenAI also exposes tokens per minute for embeddings
    ),
]
client.integrations.configure(integrations)
py docs  API docs

Additional considerations​

Data imports can be resource intensive. Consider the following when you import large amounts of data.

Asynchronous imports​

Experimental

Available starting in v1.22. This is an experimental feature. Use with caution.

To maximize import speed, enable asynchronous indexing.

To enable asynchronous indexing, set the ASYNC_INDEXING environment variable to true in your Weaviate configuration file.

weaviate:
  image: cr.weaviate.io/semitechnologies/weaviate:1.28.0
  ...
  environment:
    ASYNC_INDEXING: 'true'
  ...

Automatically add new tenants​

By default, Weaviate returns an error if you try to insert an object into a non-existent tenant. To change this behavior so Weaviate creates a new tenant, set autoTenantCreation to true in the collection definition.

The auto-tenant feature is available from v1.25.0 for batch imports, and from v1.25.2 for single object insertions as well.

Set autoTenantCreation when you create the collection, or reconfigure the collection to update the setting as needed.

Automatic tenant creation is useful when you import a large number of objects. Be cautious if your data is likely to have small inconsistencies or typos. For example, the names TenantOne, tenantOne, and TenntOne will create three different tenants.

For details, see auto-tenant.

Questions and feedback​

If you have any questions or feedback, let us know in the user forum.