DocArray 0.18 Update

DocArray is a library for nested, unstructured, multimodal data in transit, including text, image, audio, video, 3D mesh, etc. It allows deep-learning engineers to efficiently process, embed, search, recommend, store, and transfer multi-modal data with a Pythonic API.

GitHub - jina-ai/docarray: 🧬 The data structure for unstructured multimodal data · Neural Search · Vector Search · Document Store

🧬 The data structure for unstructured multimodal data · Neural Search · Vector Search · Document Store - GitHub - jina-ai/docarray: 🧬 The data structure for unstructured multimodal data · Neural Se...

GitHubjina-ai

This release contains 7 new features, 6 bug fixes, and 8 documentation improvements. It also contains breaking changes.

💥 Backwards incompatible API changes

Increased minimum versions for dependencies:

The Qdrant backend in DocArray now requires Qdrant database v0.8.0 or higher.

Other API Changes:

• The return type of DocumentArray.evaluate() changed from a single score float to a dict mapping score names to score values.
• Fully persisting data in DocumentArray using a storage backend now has to be ensured by using the context manager. Therefore, you need to wrap your write operations to a DocumentArray in a context manager like so:

my_da = DocumentArray(storage='my_storage', config=...)
with my_da:
    ...  # write operations

Alternatively, you can call the sync() method when you finish write operations:

my_da = DocumentArray(storage='my_storage', config=...)
...  # write operations
my_da.sync()

Future API Changes:

• The MongoDB-like query language syntax for filtering in the Redis backend will be deprecated soon.
• The metric and metric_name parameters in the DocumentArray.evaluate() method were renamed and accept a list type rather than a single value (as mentioned above). The old naming and type will be deprecated soon.

🆕 Features

Support geospatial filters in Redis backend (#579)

The Redis Document Store can now accept geospatial filter queries in the DocumentArray.find() method:

from docarray import Document, DocumentArray

n_dim = 3
da = DocumentArray(
    storage='redis',
    config={
        'n_dim': n_dim,
        'columns': {'location': 'geo'},
    },
)
with da:
    da.extend(
        [
            Document(id=f'r{i}', tags={'location': f"{-98.17+i},{38.71+i}"})
            for i in range(10)
        ]
    )
max_distance = 300
filter = f'@location:[-98.71 38.71 {max_distance} km] '
results = da.find(filter=filter, limit=10)
print(
    f'Locations within: {max_distance} km',
    [(doc.id, doc.tags['location']) for doc in results],
)

Results:

Locations within: 300 km [('r0', '-98.17,38.71'), ('r1', '-97.17,39.71')]

Support multiple metrics in evaluate (#643)

DocumentArray.evaluate() now supports computing evaluations for multiple metrics at once. The metric parameter is renamed to metrics, and metric_name is renamed to metric_names.

The evaluate() method expects a list for metrics and metric_names rather than a single value. For instance, instead of doing:

da2.evaluate(
    ground_truth=da1, metric='precision_at_k', metric_name='precision@k', k=5
)  # returns average_evaluation

use:

da2.evaluate(
    ground_truth=da1, metrics=['precision_at_k'], metric_names=['precision@k'], k=10
)  # returns {'precision@k': prec_at_k_average_evaluation}

The first usage will raise a deprecation warning and will be deprecated soon.

The return type is also changed: evaluate() will now return a dict mapping metric names to their average evaluation scores instead of a single score value.

For more info, check the Evaluate Matches section in the documentation.

Show server error messages in push

When using DocumentArray.push(), error messages returned by the server will show up in the stack trace. For instance, pushing a DocumentArray with a name reserved by another user will return the following error:

requests.exceptions.HTTPError: 403 Client Error: OperationNotAllowedError: Current user is not allowed to edit this artifact. Permission denied. for url: https://api.hubble.jina.ai/v2/rpc/artifact.upload

Add warnings when using MongoDB-like filter QL syntax in Redis and support native filter QL (#645)

MongoDB-like filter QL is no longer supported in the Redis backend, and this release adds support for the native Redis QL syntax. Using MongoDB-like filter QL will raise a deprecation warning and will be deprecated soon.

Therefore, instead of using:

redis_da.find(filter={'field': {'@eq': 'value'}})

use this syntax instead:

redis_da.find(filter='@field:value')

For more information, check the Redis Document Store documentation.

Add support for labeled datasets to the evaluate function (#617)

As of this release, DocumentArray.evaluate() supports labeled datasets. Labels can be added using a tag field in each Document of your DocumentArray:

import numpy as np
from docarray import Document, DocumentArray

example_da = DocumentArray([Document(tags={'label': (i % 2)}) for i in range(10)])
example_da.embeddings = np.random.random([10, 3])
example_da.match(example_da)
print(example_da.evaluate(metric='precision_at_k'))

The results of the evaluation will be stored in the evaluations field of each Document.

You can specify the label field name using the label_tag attribute:

example_da = DocumentArray(
    [Document(tags={'my_custom_label': (i % 2)}) for i in range(10)]
)
example_da.embeddings = np.random.random([10, 3])
example_da.match(example_da)
print(example_da.evaluate(metric='precision_at_k', label_tag='my_custom_label'))

Allow progress bar while batching (#628)

You can see the progress of batching documents using DocumentArray.batch() with the show_progress parameter:

import time
from docarray import Document, DocumentArray

da = DocumentArray.empty(100000)
for i in range(1, 100000):
    da.append(Document(text=str(i)))

print('append finished')

for batch in da.batch(500, show_progress=True):
    time.sleep(0.1)

Add the n_components PCA parameter to AnnLite configurations (#606)

The parameter n_components is added to AnnLite's configuration in DocArray. Use this parameter when you want to use PCA in your AnnLite backend.

🐞 Bug Fixes

Support Qdrant 0.8.0

DocArray adds support for Qdrant versions greater than or equal to v0.8.0 and drops support for previous versions. Therefore, make sure to use version 0.8.0 or higher for both qdrant-client and the Qdrant database.

Sync DocumentArray using sync() method and context manager (#625)

Fully persisting (syncing) data in a DocumentArray to a database is now ensured using either the context manager or the sync() method. Make sure to wrap write operations to a DocumentArray in a context manager like so:

my_da = DocumentArray(storage='my_storage', config=...)
with my_da:
    ...  # write operations

or use the sync() method:

my_da = DocumentArray(storage='my_storage', config=...)
...  # write operations
my_da.sync()

Close the file handler properly in load_uri_to_audio_tensor (#609)

Method load_uri_to_audio_tensor used to open a file handler without properly closing the file. This release fixes this bug and makes sure the file is opened with a context manager and is closed properly.

Fix add not performing deep copy (#582)

Concatenation operations in DocumentArray used to operate on objects in-place, without making a copy. This resulted in the following unexpected behavior:

from docarray import DocumentArray

da1 = DocumentArray.empty(3)
da2 = DocumentArray.empty(4)
da3 = DocumentArray.empty(5)
print(da1 + da2 + da3)

da1 += da2
print('length =', len(da1))  # expected length = 7 but prints length = 16

This release fixes the bug. Concatenation will operate on new copied objects each time rather than concatenating in-place.

Fix loading from a database with subindices (#581)

Prior to this release, reloading a DocumentArray configured with subindices from a database used to produce a unique ID existing error (the actual error message depends on the backend). This happened because DocumentArray attempted to index initial documents twice in the sub-index, although they had already been indexed.

This release fixes the issue.

Remove check of default value in _non_empty_fields (#565)

Serializing a Document used to ignore scores with value 0.0. For instance, the string representation of a Document might ignore the scores with value 0 and consider them as an empty field. This release fixes the issue.

📗 Documentation Improvements

• Highlight the importance of using the context manager when it comes to fully persisting data in a database. Read more in Persistence, mutations and context manager. (#613)
• Fix a mention of the convert_uri_to_datauri() method in the documentation. (#608)
• Fix the documentation build stage so that the API reference section appears correctly for Document Stores. Now you can find the API reference for Document Stores in this section. (#594)
• Fix the docstring of the set_image_normalization() method so that it mentions proper usage and aligns with PyTorch
  conventions. (#585)
• Clarify that the Query Language syntax of filter queries in DocArray depends on the Document Store used with the
  DocumentArray instance. (#586)
• Introduce a few improvements to the README example, so that the user takes into consideration the dataset size and
  requirements. (#577)
• Fix an example of plotting embeddings in the README. (#576)
• Introduce a few grammatical improvements to the What is DocArray section. (#566)

🤟 Contributors

We would like to thank all contributors to this release:

Jie Fu (@jemmyshin)
Wang Bo (@bwanglzu)
Jonathan Rowley (@jonathan-rowley)
Alex Cureton-Griffiths (@alexcg1)
Han Xiao (@hanxiao)
AlaeddineAbdessalem (@alaeddine-13)
Michael Günther (@guenthermi)
samsja (@samsja)
Johannes Messner (@JohannesMessner)
dong xiang (@dongxiang123)
Jackmin801 (@Jackmin801)
Joan Fontanals (@JoanFM)