Motivation

For Pyris V2 we need modular and reusable pipelines. This system will support different Pipeline selections in the future, which are all accessible through the same Interface. We also want to send status updates during pipeline processing.

Description

The primary objective of the pipeline subsystem is to abstract the prompting and reasoning logic associated with various LLM use cases, such as exercise creation and tutor chat, from the overall system infrastructure.

The FeaturePipelines component is the initial receiver of user queries from the Endpoint subsystem. It directs these queries to the appropriate functional units specifically designed to handle distinct use cases. These units transform user queries into structured prompts, which are subsequently directed to the LLM subsystem through the corresponding completion and chat interfaces, contingent upon the context of the query. Responses from the LLM subsystem are then conveyed back to the Endpoint subsystem for user delivery.

During query processing, the FeaturePipelines component leverages the retrieval interface from the Retrieval subsystem to obtain contextual data pertinent to the user queries. Further details about the Retrieval subsystem can be found in the later chapters of the documentation.

Simultaneously, the FeaturePipelines component communicates operational progress to the Notification Service. This service processes the progress data and interfaces with the Artemis Connector to disseminate progress notifications. In its role, the Artemis Connector forwards these notifications to the designated Artemis system. The specifics of the Artemis Connector's role and its interaction with the Artemis system are detailed in later chapters of the documentation.