Prompt Crib Sheet

Crib sheet for prompt and pipeline#


  • First create a prompt in prompt tab
  • Pin the prompt.
  • Set prompt settings - model name and max token length important.
  • Important step - Make sure to set the default integrations and default model to avoid having to set for every prompt. Make sure to use "chat/" when using Azure instance.
  • quick handy - Click outside prompt text box to save.
  • Stash the items to run the prompt against
  • Different ways of including stashed data into prompts
    • {{ sourceConversation }} - does source conversation stay the same.
    • {{ conversation }} - includes the entire conversation
    • {{ text }} - includes utterances
  • Test the pinned prompt against the stashed items.
  • Output of every prompt run can be accessed using
    • promptId
    • generationRunId
  • In case if you are not happy with your results, modify the prompt to get required results (prompt-tuning).
  • Once happy with the results, set up your pipeline.


  • Go to pipeline tab and create a new pipeline.
  • Pin the pipeline to edit it.
  • Editing the name of the pipeline, selecting prompt and choosing the number of items to be processed can be edited on the pinned pipeline whereas input data, filters, NLU engine, and sort by must be decided on the data tab upon clicking the edit button.
  • Save and run pipeline. Now the prompt is run against the given number of data upon executing the pipeline.
  • Output of every pipeline run can be accessed using
    • pipelineId
    • pipelineStepId
  • Pipeline cache โ€“ Pipeline uses the same output if there is no modification made to the prompt, prompt settings and pipeline settings.
  • Error handling - If the model throws an error, you would see the original utterance with an error metadata.

Pipeline use cases#

  • topic analysis
  • model building from scratch
  • getting resolved conversations
  • finding call driver/key issues
  • identifying the actions taken by the agent
  • multi-pipeline runs (pipeline on pipeline), etc.