AI Power: Natural Language to Structured Data Query

Unlock the true power of prompt engineering by directly leveraging recently enhanced LLM capabilities and libraries like `litellm` to output validated Pydantic models. This ensures your human intent immediately transforms into machine-ready, structured data, skipping the messy parsing.

python
from pydantic import BaseModel, Field
from enum import Enum
import litellm # A recent update to LiteLLM or its underlying models improves Pydantic reliability

# Define your desired structured data shape with Pydantic
class AggregateFunction(str, Enum):
    SUM = "sum"
    MEAN = "mean"
    COUNT = "count"
    MAX = "max"
    MIN = "min"

class QueryParameter(BaseModel):
    column: str = Field(description="The column name to apply the operation on.")
    value: str | int | float | None = Field(default=None, description="Value for filtering or comparison.")
    operation: str | None = Field(default=None, description="Operation like 'equals', 'greater_than', 'contains'.")
    aggregate: AggregateFunction | None = Field(default=None, description="Aggregation function to apply.")

class DataQuery(BaseModel):
    select_columns: list[str] = Field(description="List of columns to select.")
    filters: list[QueryParameter] = Field(default_factory=list, description="List of filter conditions.")
    aggregations: list[QueryParameter] = Field(default_factory=list, description="List of aggregation operations.")
    group_by_columns: list[str] = Field(default_factory=list, description="Columns to group by for aggregations.")

# The human's high-level request
user_request = "I need the average 'sales' and total 'profit' for 'region'='East' and 'product_category' containing 'Electronics', grouped by 'region' and 'month'."

# 🧠 The AI-augmented workflow: direct Pydantic model generation using Llama 3.1
# Llama 3.1 (released recently) excels at structured output when guided well.
response_object: DataQuery = litellm.completion(
    model="ollama/llama3.1", # Use a powerful, recent SLM or cloud model
    messages=[
        {"role": "system", "content": "You are an expert data query generator. Based on the user's request, produce a structured data query adhering to the provided Pydantic schema for analysis with Polars/DuckDB. Ensure all fields are correctly populated and types match perfectly. Be concise."}, # Clear instructions
        {"role": "user", "content": f"Generate a DataQuery object for: '{user_request}'"}
    ],
    response_model=DataQuery, # LiteLLM's powerful Pydantic integration ensures valid output
    temperature=0.1 # Keep it low for structured tasks
).choices[0].message.content # LiteLLM handles the parsing and validation for you!

print(response_object.model_dump_json(indent=2))