Skip to content

Prompt Details

The prompt is the instruction that shapes how an AI Processor behaves. Every AI Processor — whether it is part of a Pack Pipeline, a Change Instance Validator, or any other type — requires a prompt that tells the LLM what to do with the incoming data.

The prompt you define is the serviceowner_prompt in the final payload sent to the LLM. It is combined with the system-level prompt from the LLM Model (netorca_prompt) and the contextual data (declaration, PackData from previous stages, documents, etc.) based on the enabled options.

Writing Effective Prompts

A good prompt should:

  • Be specific about the expected output format (JSON, markdown, structured report)
  • Clearly state the task the LLM should perform
  • Reference the type of data the LLM will receive (e.g., service item declarations, previous configurations)
  • Include any constraints or rules the output must follow (naming conventions, security policies, vendor-specific requirements)

The more specific your prompt, the more accurate and consistent the AI Processor's output will be. Combine prompts with Documents (Knowledge) to provide domain-specific context that improves output quality without making the prompt itself overly long.

How the Prompt is Used

The Service Owner's prompt is one part of the final payload sent to the LLM. NetOrca assembles the full prompt by combining:

  1. netorca_prompt — The system-level prompt defined in the LLM Model, acting as a wrapper for all AI interactions.
  2. serviceowner_prompt — The prompt you define on the AI Processor (this page).
  3. Contextual data — Service item declaration, PackData from previous stages, change instance details, documents, and other data depending on the enabled options.

For the full payload structure, see Pack Pipeline — Final Prompt and Change Instance Validator — Final Prompt.

Tips

  • Iterate on your prompts — Use the retrigger mechanism to test and refine prompts based on actual outputs.
  • Use Documents for static knowledge — Move vendor guides, naming conventions, and policy documents into Documents (Knowledge) rather than embedding them directly in the prompt.
  • Use Response Schema for structure — If you need a specific output format, define a Response JSON Schema rather than relying on prompt instructions alone.
  • Use Generative UI for human-readable output — If the output is intended for users in the GUI, enable Generative UI and instruct the prompt to return structured UI components.