vertex-batch 0.1.19

A module for batch processing with Google Cloud Storage and MongoDB integration.

vertex_batch Library Documentation

vertex_batch is a Python library designed for LLM google vertex batch processing. It provides abstractions for database operations, batch file handling, and callback server management, enabling scalable and robust batch processing pipelines.

Features

• Batch Database Management: Store, update, and retrieve conversational payloads for batch processing.
• Batch File Generation: Write payloads to batch files for LLM processing (e.g., Gemini).
• Callback Server: Receive and process batch results via HTTP callbacks.
• Integration with LLM Workflows: Easily integrate with Celery tasks and main application logic.

Installation

Add vertex_batch to your project (ensure it's in your Python path):

pip install vertex_batch

Usage

1. Database Operations

Use vertex_batch.db.Db to interact with the batch database.

from vertex_batch.db import Db

db = Db(
    url="...",
    db_name="...",
    batch_collection_name="..."
)

# Save, update, and retrieve payloads
db.save_payload(payload)
db.update_payload(custom_id="...", status="DONE")
payloads = db.get_payloads(status="PENDING")

2. Batch File Generation

Use vertex_batch.file.File to write payloads to batch files and process them with LLMs.

from vertex_batch.file import File
from pathlib import Path
from datetime import datetime

file = File(
    db=db,
    folder_path=Path("batchs_files/input"),
    file_name_format=f"voc_batch_{datetime.now().strftime('%Y%m%d_%H%M%S')}_batch.jsonl",
    gemini_model="publishers/google/models/gemini-2.5-flash"
)

payloads = db.get_payloads(status="PENDING")
file.write(paylods=payloads)
file.process()

3. Line Management

Use vertex_batch.line.Line to save individual batch lines for LLM processing.

from vertex_batch.line import Line

line = Line(
    db=db,
    custom_id="conversationId-vocId",
    user_prompt="User prompt text",
    sys_prompt="System prompt text",
    **kwargs
)
line.save()

4. Callback Server

Use vertex_batch.callback.Callback to start a callback server that receives batch results and processes them.

from vertex_batch.callback import Callback
from pathlib import Path

def treat_answers(payloads, file_path):
    # Custom logic to process batch results
    pass

callback = Callback(
    db=db,
    port=8010,
    destination_dir=Path("batchs_files/output"),
    func=treat_answers
)

callback.start_server()

You can run the callback server in a separate thread:

import threading

threading.Thread(target=callback.start_server, daemon=True).start()

NOTE : the callback path is /batch_processing_done

Example Workflow

1. Save payloads to the batch database using Db.
2. Generate batch files for LLM processing using File.
3. Process batch results via the callback server (Callback), which calls your custom handler (treat_answers).
4. Update payloads and send results to downstream systems (e.g., Kafka).

Integration with Celery

You can use vertex_batch in Celery tasks for asynchronous batch processing:

from vertex_batch.db import Db
from vertex_batch.line import Line
from vertex_batch.file import File

@celery_app.task(name="...")
def function_task():
    db = Db(...)
    file = File(...)
    payloads = db.get_payloads(status="PENDING")
    file.write(paylods=payloads)
    file.process()

API Reference

Db

• Db.save_payload(payload)
• Db.update_payload(custom_id, status=None, answer=None)
• Db.get_payloads(status)
• Db.get_payload(custom_id)
• Db.flag_payloads(file_path, flag)

File

• File.write(paylods, is_relaunch=False)
• File.process()

Line

• Line.save()

Callback

• Callback(db, port, destination_dir, func)
• Callback.start_server()

DO NOT FORGET TO SET THOSE OS ENV VARIABLES

• GOOGLE_APPLICATION_CREDENTIALS
• GOOGLE_STORAGE_BUCKET
• GOOGLE_PROJECT_NAME
• GOOGLE_PROJECT_LOCATION
• BATCH_FILE_SIZE_LIMIT
• MONGO_DB_URL

License

MIT License : AYOUB ERRKHIS

Contributing

Contributions are welcome! Please submit issues or pull requests via GitHub.

Contact

For support or questions, contact the maintainers or open an issue on the repository.