piglets 0.1.8

A modular text-to-SQL toolkit.

🐷 piglets

A modular library of text-to-SQL tools.

Status

piglets is currently an alpha-stage package. The API is expected to evolve before 1.0.

Get started

Install

venv

pip install piglets

uv

uv add piglets

Install the optional dependency for the model provider you use. For OpenAI:

venv

pip install "piglets[openai]"

uv

uv add "piglets[openai]"

Other provider extras include anthropic, google_genai, google_vertexai, bedrock, cohere, mistralai, groq, ollama, and openrouter.

Install the optional dependency for the database backend you use. For BigQuery:

venv

pip install "piglets[bigquery]"

uv

uv add "piglets[bigquery]"

Logical planning

Use gpt-5.2 to generate 3 logical plans from a natural language query.

from piglets import LogicalPlanner

# initialise a logical planner
logical_planner = LogicalPlanner('gpt-5.2')

# generate 3 logical plan samples and aggregate them
logical_plan = logical_planner.plan(
    natural_language_query="What was the average number of piglets per week for Q4 2025?",
    num_samples=3,
)

# print the aggregated logical plan
for i, step in enumerate(logical_plan.logical_steps):
    print(f"Step {i + 1}: ")
    print(step)

# inspect the candidate plans used to create the aggregate
print(f"Aggregated from {len(logical_plan.sample_plans)} sample plans.")

>>> Step 1:
>>> 1. Identify all piglet birth (or piglet addition) events with their event dates and piglet counts.
>>> Step 2:
>>> 2. Filter the events to the Q4 2025 date range (Oct 1, 2025 through Dec 31, 2025).
>>> Step 3:
>>> 3. Assign each event to a calendar week within that quarter using a consistent week definition (e.g., week starting Monday or Sunday).
>>> Aggregated from 3 sample plans.
...

Database connector

Use DatabaseConnector to inspect a supported database and return a typed schema.

from piglets import DatabaseConnector

database_connector = DatabaseConnector(
    database_type="bigquery",
    database_name="my_bigquery_dataset",
)

database = database_connector.get_database_schema()

print(database.name)
for table in database.tables:
    print(table.name)
    for column in table.columns:
        print(f"- {column.name} ({column.data_type})")

BigQuery connections use the GOOGLE_CLOUD_PROJECT_ID environment variable by default. You can also pass gcp_project_id directly:

database_connector = DatabaseConnector(
    database_type="bigquery",
    database_name="my_bigquery_dataset",
    gcp_project_id="my-gcp-project",
)

Dual-pathway pruning

Use Pruner to reduce a database schema with both preservation and deletion signals. The preservation pathway selects tables and columns that look useful for the query. The deletion pathway removes tables and columns that look irrelevant. dual_pathway_pruning() combines both paths into a final Database schema.

from piglets import DatabaseConnector, LogicalPlanner, Pruner

question = "Which tags saw the largest increase in average answer score from 2022 to 2023, considering only questions with at least 5 answers?"

logical_planner = LogicalPlanner("gpt-5.2")
logical_plan = logical_planner.plan(
    natural_language_query=question,
    num_samples=3,
)

database_connector = DatabaseConnector(
    database_type="bigquery",
    database_name="stack_overflow",
)
database = database_connector.get_database_schema()

pruner = Pruner(model_name="gpt-5.2")
pruned_database = pruner.dual_pathway_pruning(
    natural_language_query=question,
    database=database,
    logical_plan=logical_plan,
)

print(pruned_database.export_as_string())

Current scope

Database

DatabaseConnector currently supports BigQuery. It connects to a database by database_name and returns a Database object containing Table and Column objects.

Planning

The first included primitive is a LogicalPlanner that turns a natural-language analytics question into an ordered list of abstract logical steps. The logical planner is an implementation of the planner found in the Apex-SQL paper here.

The LogicalPlanner has a plan method that can generate one plan or sample multiple plans and aggregate them with num_samples.

Plan aggregation is available through LogicalPlans.aggregate(). Aggregated plans include a sample_plans attribute containing the candidate LogicalPlan objects used to produce the final plan.

Pruning

Pruner supports preservation pruning, deletion pruning, and dual-pathway pruning. Preservation pruning returns a PreservationSet of useful tables and columns. Deletion pruning returns a DeletionSet of irrelevant tables and columns. Dual-pathway pruning combines both into a final pruned Database.