Skip to content

15. PromptFlow: Codebase

[!NOTE] Our environment, resources and connections are configured. Now, let's learn about prompt flow and how it works. A prompt flow is a DAG (directed acyclic graph) made of up nodes connected together in a flow. Each node is a function tool (written in Python) that can be edited and customized to suit your needs.

  • [] 01 | Let's explore the Prompt Flow extension

    • Click the "Prompt Flow" icon in the Visual Studio Code sidebar
    • You should see a slide-out menu with the following sections
      • Quick Access - Create new flows, install dependencies etc,
      • Flows - Lists flows in project (defined by flow.dag.yaml)
      • Tools - Lists available function tools (used in flow nodes)
      • Batch Run History - flows run against data or other runs
      • Connections - Lists connections & helps create them
    • We'll revisit this later as needed, when executing prompt flows.

  • [] 02 | Let's understand prompt flow folders & structure

    • Click the "Explorer" icon in the Visual Studio Code sidebar
    • Promptflow can create three kinds of flows:
      • standard = basic flow folder structure
      • chat = enhances standard flow for conversations
      • evaluation = special flow, assesses outputs of other flows
    • Explore the "contoso_chat" folder for a chat flow:
      • flow.dag.yaml - defines the flow (inputs, outputs, nodes)
      • source code (.py, .jinja2) - function tools used by flow
      • requirements.txt - defines Python dependencies for flow
    • Explore the "eval/" folder for examples of eval flows
      • eval/groundedness - tests for single metric (groundedness)
      • eval/multi_flow - tests for multiple metrics (groundedness, fluency, coherance, relevance)
      • eval/evaluate-chat-prompt-flow.ipynb - shows how these are used to evaluate the_contoso_chat_ flow.

  • [] 03 | Let's explore a prompt flow in code

    • Open Visual Studio Code file: contoso-chat/flow.dag.yaml
    • You should see a declarative file with these sections:
      • environment - requirements.txt to install dependencies
      • inputs - named inputs & properties for flow
      • outputs - named outputs & properties for flow
      • nodes - processing functions (tools) for workflow

The "prompt flow" is defined by the flow.dag.yaml but the text view does not help us understand the "flow" of this process. Thankfully, the Prompt Flow extension gives us a Visual Editor that can help. Let's explore it.

🥳 Congratulations!
You're ready to explore a prompt flow visually!