Model Prompts#

Some Agents employ structured Pydantic models as their response format so that the code easily use the responses. These models can also be overridden by specifying the model key in prompts.

For example, the SQLAgent’s main default model is:

class Sql(BaseModel):

    chain_of_thought: str = Field(
        description="""
        You are a world-class SQL expert, and your fame is on the line so don't mess up.
        Then, think step by step on how you might approach this problem in an optimal way.
        If it's simple, just provide one sentence.
        """
    )

    expr_slug: str = Field(
        description="""
        Give the SQL expression a concise, but descriptive, slug that includes whatever transforms were applied to it,
        e.g. top_5_athletes_gold_medals
        """
    )

    query: str = Field(description="Expertly optimized, valid SQL query to be executed; do NOT add extraneous comments.")

To override the chain_of_thought field, you can subclass the Sql model:

from lumen.ai.models import Sql

class CustomSql(Sql):
    chain_of_thought: str = Field(
        description="Think through the query like an expert DuckDB user."
    )

Then, you can specify the model key in prompts:

prompts = {
    "main": {
        "model": CustomSql
    }
}
agents = [lmai.agents.SQLAgent(prompts=prompts)]
ui = lmai.ExplorerUI(agents=agents)

Note, the field names in the model must match the original model’s field names, or else unexpected fields will not be used.

This Page