DataChain

DataChain is a Python-based AI-data warehouse for transforming and analyzing unstructured data like images, audio, videos, text and PDFs. It integrates with external storage (e.g. S3) to process data efficiently without data duplication and manages metadata in an internal database for easy and efficient querying.

Use Cases

1. ETL. Pythonic framework for describing and running unstructured data transformations and enrichments, applying models to data, including LLMs.
2. Analytics. DataChain dataset is a table that combines all the information about data objects in one place + it provides dataframe-like API and vecrorized engine to do analytics on these tables at scale.
3. Versioning. DataChain doesn't store, require moving or copying data (unlike DVC). Perfect use case is a bucket with thousands or millions of images, videos, audio, PDFs.

Getting Started

Visit Quick Start and Docs to get started with DataChain and learn more.

pip install datachain

Example: download subset of files based on metadata

Sometimes users only need to download a specific subset of files from cloud storage, rather than the entire dataset. For example, you could use a JSON file's metadata to download just cat images with high confidence scores.

from datachain import Column, DataChain

meta = DataChain.from_json("gs://datachain-demo/dogs-and-cats/*json", object_name="meta", anon=True)
images = DataChain.from_storage("gs://datachain-demo/dogs-and-cats/*jpg", anon=True)

images_id = images.map(id=lambda file: file.path.split('.')[-2])
annotated = images_id.merge(meta, on="id", right_on="meta.id")

likely_cats = annotated.filter((Column("meta.inference.confidence") > 0.93) \
                               & (Column("meta.inference.class_") == "cat"))
likely_cats.export_files("high-confidence-cats/", signal="file")

Example: LLM based text-file evaluation

In this example, we evaluate chatbot conversations stored in text files using LLM based evaluation.

$ pip install mistralai # Requires version >=1.0.0
$ export MISTRAL_API_KEY=_your_key_

Python code:

from mistralai import Mistral
from datachain import File, DataChain, Column

PROMPT = "Was this dialog successful? Answer in a single word: Success or Failure."

def eval_dialogue(file: File) -> bool:
     client = Mistral()
     response = client.chat.complete(
         model="open-mixtral-8x22b",
         messages=[{"role": "system", "content": PROMPT},
                   {"role": "user", "content": file.read()}])
     result = response.choices[0].message.content
     return result.lower().startswith("success")

chain = (
   DataChain.from_storage("gs://datachain-demo/chatbot-KiT/", object_name="file", anon=True)
   .settings(parallel=4, cache=True)
   .map(is_success=eval_dialogue)
   .save("mistral_files")
)

successful_chain = chain.filter(Column("is_success") == True)
successful_chain.export_files("./output_mistral")

print(f"{successful_chain.count()} files were exported")

With the instruction above, the Mistral model considers 31/50 files to hold the successful dialogues:

$ ls output_mistral/datachain-demo/chatbot-KiT/
1.txt  15.txt 18.txt 2.txt  22.txt 25.txt 28.txt 33.txt 37.txt 4.txt  41.txt ...
$ ls output_mistral/datachain-demo/chatbot-KiT/ | wc -l
31

Key Features

📂 Multimodal Dataset Versioning.

• Version unstructured data without moving or creating data copies, by supporting references to S3, GCP, Azure, and local file systems.
• Multimodal data support: images, video, text, PDFs, JSONs, CSVs, parquet, etc.
• Unite files and metadata together into persistent, versioned, columnar datasets.

🐍 Python-friendly.

• Operate on Python objects and object fields: float scores, strings, matrixes, LLM response objects.
• Run Python code in a high-scale, terabytes size datasets, with built-in parallelization and memory-efficient computing — no SQL or Spark required.

🧠 Data Enrichment and Processing.

• Generate metadata using local AI models and LLM APIs.
• Filter, join, and group datasets by metadata. Search by vector embeddings.
• High-performance vectorized operations on Python objects: sum, count, avg, etc.
• Pass datasets to Pytorch and Tensorflow, or export them back into storage.

Contributing

Contributions are very welcome. To learn more, see the Contributor Guide.

Community and Support

• Docs
• File an issue if you encounter any problems
• Discord Chat
• Email
• Twitter

DataChain Studio Platform

DataChain Studio is a proprietary solution for teams that offers:

• Centralized dataset registry to manage data, code and dependency dependencies in one place.
• Data Lineage for data sources as well as derivative dataset.
• UI for Multimodal Data like images, videos, and PDFs.
• Scalable Compute to handle large datasets (100M+ files) and in-house AI model inference.
• Access control including SSO and team based collaboration.