This repository contains a Google Colab or Jupyter notebook that shows how to use a large language model (LLM) to analyze period‑over‑period trends in your own data. In the included example, we want to know: how did health insurance claims change from one period (early 2023) to another period (early 2024), and what were the underlying drivers? We need two data cubes as input: one data cube contains the "descriptor" data (claims, with all imaginable drivers loaded into columns), the other data cube contains the "normalizer" data (membership, with all imaginable membership segments loaded into columns).

For example, claims could be up 20% year-over-year, but membership overall is up 10%, so the per-member-per-month (PMPM) claims trend is only 10%. Or, FL claims last year could have run at $100 PMPM, and TX claims at $50 PMPM. None of that changed, but the membership mix went from 50%/50% FL/TX to 60%/40% FL/TX, which means that simply due to that mix shift, total claims PMPM went from $75 to $80 PMPM.

The notebook defines an LLM-based agent that will repeatedly query and combine these two data cubes to create an attribution or decomposition of the trend you point it at.

The notebook lets you:

1. Build BigQuery tables from a YAML configuration describing dimensions, metrics and source tables.
2. Define a system prompt and analysis plan for a conversational agent.
3. Expose data‑access functions (SQL queries, dimension lookups, chart creation) as tools the agent can call.
4. Let the agent iteratively drill into the data, write findings to a Google Doc and generate charts.

1. Open the notebook in Colab or Jupyter and install the required Python packages when prompted.
2. Provide your Google Cloud project IDs and authentication credentials so the BigQuery queries run correctly.
3. Modify dimensions.yml or the YAML cell in the notebook to point to your own claim and membership tables and to customize dimensions or metrics.
4. Execute the cells to build the normalized data tables and launch the agent. The agent will produce a step‑by‑step analysis of the 2023→2024 trend and write a summary to a Google Doc.

• Adapt the YAML configuration to describe the dimensions, metrics and tables in your data warehouse.
• Customize the analysis plan in the notebook to specify the question you want the LLM to answer. The default plan walks through a pyramid‑style drill‑down.
• Add or remove tool wrappers to expose whatever data queries or visualization functions are useful for your analysis.
• Run the notebook and monitor the agent as it iteratively queries your data and records findings.

Important: This isa proof of concept and may produce inaccurate or incomplete interpretations. Always review the generated analysis before relying on it.