PromptFlow is an open-source development framework created by Microsoft for building, evaluating, and deploying LLM-based applications through executable workflows called flows.¹⁾ It connects LLMs, prompts, Python code, and external tools into directed acyclic graphs (DAGs) that can be tested, debugged, and iterated systematically. PromptFlow integrates tightly with Azure Machine Learning and Microsoft Foundry.

• GitHub: microsoft/promptflow|github.com/microsoft/promptflow]]
• Available as: VS Code extension, Python SDK, CLI, Azure portal
• License: MIT

Flows are executable workflows represented as DAGs where each node is one of:

• LLM Node: Calls a language model (GPT-4, Codex, etc.) with a prompt template
• Python Node: Runs custom Python code for data transformation or business logic
• Tool Node: Invokes external APIs or services
• Prompt Node: Manages prompt templates with variable substitution

Flows chain these nodes together, passing outputs from one as inputs to the next, creating reproducible pipelines for LLM application logic.

PromptFlow supports a structured development lifecycle:²⁾

1. Initialization: Define business use case, create basic flow with templates
2. Experimentation: Test flow on sample data, debug node-by-node with visual tools
3. Evaluation: Run batch evaluations on large datasets with built-in quality metrics
4. Refinement: Iterate on prompts, model selection, and parameters based on evaluation results
5. Production: Optimize, deploy, and monitor in production environment

The visual DAG editor makes complex prompt engineering accessible without deep coding expertise, while the SDK supports programmatic flow construction for advanced users.

PromptFlow excels in systematic LLM application testing:³⁾

• Batch evaluation: Run flows against entire datasets to measure performance
• Built-in metrics: Quality, relevance, groundedness, fluency scores
• A/B testing: Compare different prompt versions or model configurations
• Prompt tuning: Systematic optimization of prompt templates
• Experiment tracking: Version and compare evaluation results over time

In Microsoft Foundry (classic), additional features include advanced quality assessments and comprehensive tracking dashboards.

PromptFlow is deeply embedded in the Microsoft Azure ecosystem:⁴⁾

• Azure Machine Learning Studio: Visual flow editor, managed compute, deployment
• Microsoft Foundry (classic): Enhanced evaluation, blue/green deployments, monitoring
• CI/CD Pipelines: Direct integration with Azure DevOps and GitHub Actions
• Managed Endpoints: One-click deployment to Azure-hosted inference endpoints

The cloud version handles backend complexity (compute, scaling, monitoring) while maintaining the same flow definitions used locally.

• VS Code Extension: Full visual editor with flow debugging and testing
• Python SDK: Programmatic flow construction, evaluation, and deployment
• CLI: Command-line tools for automation and CI/CD integration
• Run identical flows locally and in the cloud

from promptflow.core import tool, Prompty
from promptflow.client import PFClient
 
# Define individual tools (nodes) for the flow
@tool
def extract_keywords(text: str) -> liststr:
    """Extract key topics from input text."""
    stop_words = {"the", "a", "is", "in", "to", "and", "of", "for", "on", "with"}
    words = text.lower().split()
    return [w for w in words if len(w) > 3 and w not in stop_words]
 
@tool
def classify_intent(keywords: liststr) -> str:
    """Classify user intent based on extracted keywords."""
    intents = {
        "technical": ["code", "error", "debug", "api", "function", "deploy"],
        "billing": ["price", "cost", "subscription", "payment", "invoice"],
        "general": ["help", "question", "info", "about", "learn"],
    }
    for intent, triggers in intents.items():
        if any(kw in triggers for kw in keywords):
            return intent
    return "general"
 
@tool
def route_response(intent: str, original_text: str) -> dict:
    """Route to appropriate handler based on classified intent."""
    handlers = {
        "technical": "Routing to technical support team...",
        "billing": "Routing to billing department...",
        "general": "Routing to general assistant...",
    }
    return {
        "intent": intent,
        "handler": handlers.get(intent, handlers["general"]),
        "original_query": original_text,
    }
 
# Execute the DAG: extract_keywords -> classify_intent -> route_response
user_input = "I need help debugging an API error in my deployment"
keywords = extract_keywords(user_input)
print(f"Keywords: {keywords}")
 
intent = classify_intent(keywords)
print(f"Intent: {intent}")
 
result = route_response(intent, user_input)
print(f"Routing: {result}")
 
# To run as a full PromptFlow flow, define flow.dag.yaml:
flow_yaml = """
inputs:
  text:
    type: string
nodes:
  - name: extract_keywords
    type: python
    source: {type: code, path: flow_tools.py}
    inputs: {text: ${inputs.text}}
  - name: classify_intent
    type: python
    source: {type: code, path: flow_tools.py}
    inputs: {keywords: ${extract_keywords.output}}
  - name: route_response
    type: python
    source: {type: code, path: flow_tools.py}
    inputs: {intent: ${classify_intent.output}, original_text: ${inputs.text}}
outputs:
  result:
    value: ${route_response.output}
"""
print(f"\nFlow DAG YAML:\n{flow_yaml}")

• Promptfoo
• Prompt Chaining
• Flowise
• Prompt Requests
• System Prompt Architecture

• LangChain
• LlamaIndex
• Tool Integration Patterns
• Retrieval-Augmented Generation