Natural Language Delete¶

NLQueryDeleterExecutor deletes records matching natural language descriptions. It includes a preview step to avoid accidental deletions.

Basic Usage¶

from amsdal_ml.ml_models.openai_model import OpenAIModel
from amsdal_ml.ml_retrievers.query_deleter import NLQueryDeleterExecutor

llm = OpenAIModel()
llm.setup()

deleter = NLQueryDeleterExecutor(llm=llm, queryset=Customer.objects.all())

analysis = await deleter.analyze('Delete inactive customers from 2023')
print(analysis.total)     # number of records to delete

deleted = await deleter.execute(analysis)

How It Works¶

The delete operation follows a 2-phase pipeline:

Phase 1: Split¶

The LLM derives a filter query from the delete instruction (e.g., "Delete inactive customers from 2023" → filter by is_active=False and created_at in 2023).

Phase 2: Analysis & Execution¶

The executor:

Runs the filter query to find matching records
Returns a preview of affected records
Executes deletion within a transaction

Preview Before Delete¶

Use analyze() to preview which records will be deleted:

deleter = NLQueryDeleterExecutor(llm=llm, queryset=Customer.objects.all())

analysis = await deleter.analyze('Delete all test customers')

print(f'Records to delete: {analysis.total}')
for record in analysis.records:
    print(f'  {record}')

# Only execute if confirmed
if confirmed:
    deleted = await deleter.execute(analysis)

Authorization¶

Pass a callback to check permissions before deleting:

deleter = NLQueryDeleterExecutor(
    llm=llm,
    queryset=Customer.objects.all(),
    on_before_delete=my_auth_callback,
)

Document-Oriented Facade¶

from amsdal_ml.ml_retrievers.query_deleter import NLQueryDeleter

deleter = NLQueryDeleter(llm=llm, queryset=Customer.objects.all())
analysis = await deleter.analyze('Remove archived customers')
documents = await deleter.invoke(analysis)