Overview¶

This document discuss the architecture of the solution.

Architecture¶

The following diagram illustrates current architecture.

Widget calls to Data Access API to fetch data.

1.1 (Optional) Data Access API calls to Gemini to generate follow up questions, based on the user's profile and dataset.
Widget renders the dataset and questions.
User double click into the data point.
Widget sends the specific data point and the question to Agent API.
Agent use tools to fetch relevant data when required, and generate an answer to the question.
Agent API repond with the final results.
Widget renders the results.

Widget calls to Agent API with a pre-defined question.
Agent API generates a SQL query based on the user's profile and the question.
Agent Use tools to fetch relevant data when required, and generate follow up questions.
Agent API repond with the final results.
Widget renders the dataset and follow up questions..
User double click into the data point.
Widget sends the specific data point and the question to Agent API.
Agent use tools to fetch relevant data when required, and generate an answer to the question.
Agent API repond with the final results.
Widget renders the results.

	NL to SQL	Pre-scripted SQL / Data API
Use when...	You want to provide your user a free-form query experience	You already have an API or SQL for you report, and you will use it as a start
	The task is expected to be time-consuming, for example, the task is to analyze a large dataset	The task requires fast data retrival.

	Large model	Small Model
Example	Gemini-2.5-Pro	Gemini-2.0-Flash
Use when	The task requires complex reasoning.	The task does not requires complex reasoning
Example	Analyze a given dataset and provide follow up actions	Generate questions based on user's profile