The Nebuly SDK is a Python-based library, meticulously crafted to integrate seamlessly into your application. This document offers a comprehensive overview of the SDK’s architecture and its functionality within your application.

​
Architecture Diagram

Alt Text

​
Architecture Components

​
Your Application

This is the environment where the SDK is installed and operates. It integrates into the primary thread of your application.

​
Nebuly Observer

The Observer is a component that monitors the calls to various LLM providers. It gathers the input and output of these calls, collectively referred to as an “interaction”. An interaction may encompass multiple calls to the same or different LLM providers, termed as “spans”.

​
Interaction Queue

The Interaction Queue serves as a temporary storage for interactions. It retains the interactions until they are ready for processing by the Interaction Parser. This mechanism decouples the Primary Thread from the Nebuly Thread, enabling concurrent operation.

​
Interaction Parser

The Interaction Parser is tasked with preparing the interactions for transmission to the Nebuly Server. It operates on a separate Nebuly Thread, ensuring that the primary application thread remains unblocked. The Parser is equipped to capture input and output based on the specific LLM providers.

​
Data Publisher

The Data Publisher component transmits the parsed interactions to the Nebuly Server. It ensures reliable and efficient data transmission.

​
Nebuly Server

The Nebuly Server is the destination where the interactions are received and processed. It calculates various metrics such as user satisfaction, cost, topic, frustration level, and latency.

​
Performance Impact

The Nebuly SDK is engineered to have a minimal impact on your application’s performance. It operates on a separate thread, ensuring that your primary application thread remains unblocked. In the event of any issues, the SDK is capable of restarting its thread, ensuring uninterrupted operation.

​
Error Handling

The SDK is designed to be robust and resilient to errors. It includes try-except blocks to handle any potential issues and ensure smooth operation.