dataset: gross receipt taxes of New Mexico time period: 2009 - 2023

The project premise is to fine tune an open source LLM like Llama 2 with a specific dataset. The goal is to train the model with natural language prompts in order to generate sql queries to our database. We would set a "correct answers returned" success rate we would want to achieve (i.e. > 85%). Essentially making an assistant for generating queries for the data.

This tool would enable users to quickly understand features of the data like revenue generation, sector performance, and broader economic patterns at both the state and local levels; potentially becoming an invaluable asset for revenue forecasting, policy formation, and economic analytics.

Further work on this project includes predictive analytics for state tax data.

A locally hosted Llama 2 model is fine-tuned on a custom dataset of natural language questions paired with their corresponding SQL queries. Training data is generated from the database schema and covers queries across regions, industries, tax elements, and date ranges.

LangChain is used with a local LlamaCpp model to create a SQL database chain. A custom prompt template provides the database schema as context, enabling the LLM to generate and execute SQL queries against the database from natural language input.

LlamaIndex provides a NLSQLTableQueryEngine that connects directly to the database and translates natural language questions into SQL queries using an LLM, with built-in token counting for cost tracking.

1. Copy .env.example to .env and fill in your database credentials and API keys:

   cp .env.example .env
   

2. Install dependencies:

   pip install python-dotenv sqlalchemy pymysql llama-index langchain langchain-experimental
   

What counties are included in the tax data?

SELECT *
FROM entity
WHERE id IN (
        SELECT entity_id
        FROM (
                SELECT *
                FROM
                    entity_category_rel ecr
                    JOIN entity_category ec on ecr.category_id = ec.id
            ) AS categories
        WHERE
            categories.name = "County"
    );

What was the total retail trade industry tax paid for Albuquerque in April 2018

SELECT total_tax_paid
FROM industry_entity_tax
WHERE report_id IN (
        SELECT id
        FROM report
        WHERE
            reporting_month = "2018-04-01"
    )
    AND industry_entity_id IN (
        SELECT id
        FROM industry_entity
        WHERE entity_id IN (
                SELECT id
                FROM entity
                WHERE
                    name LIKE "%Albuquerque%"
            )
            AND `name` LIKE "%Retail Trade%"
    )
    AND `name` = "Total";

What was the maximum construction industry taxable receipts for Bernalillo County in 2016?

SELECT MAX(taxable_receipts) as max_taxable_receipts
FROM industry_entity_tax
WHERE report_id IN (
        SELECT id
        FROM report
        WHERE
            reporting_month BETWEEN "2016-01-01" AND "2016-12-31"
    )
    AND industry_entity_id IN (
        SELECT id
        FROM industry_entity
        WHERE entity_id IN (
                SELECT id
                FROM entity
                WHERE
                    name LIKE "%Bernalillo County%"
            )
            AND `name` LIKE "%Construction%"
    )
    AND `name` = "Total";