microsoft/promptflow

Prompt flow – Overview

What it is – Prompt flow is a Microsoft‑maintained toolkit for building, testing, and deploying applications that use large language models (LLMs). It ties together prompts, Python code, and other “tools” into a single flow that can be run locally, in CI/CD pipelines, or as a cloud service.

Key capabilities

• Flow authoring – Define a DAG (flow.dag.yaml) that connects LLM calls, prompts, and custom Python nodes. A CLI (pf) can scaffold a new flow (e.g., a chat bot) and manage connections to OpenAI or Azure OpenAI.
• Debug & tracing – Interactive testing (pf flow test --interactive) lets you step through a flow and see each LLM interaction, making prompt‑engineering iterative.
• Evaluation – Run a flow over larger datasets, collect metrics, and plug the results into CI/CD to enforce quality before release.
• Production readiness – Deploy the flow to any serving platform, or embed it directly in an application. Azure AI offers a managed cloud version.
• IDE support – A VS Code extension provides a visual flow designer.

Typical workflow

1. Create a flow (pf flow init …).
2. Configure a connection for your LLM API key (pf connection create …).
3. Develop prompts and Python nodes, testing interactively (pf flow test).
4. Evaluate on test sets, integrate with CI/CD.
5. Deploy to Azure or your own runtime.

Status – The project is slated for retirement on 20 April 2027; Microsoft recommends migrating new work to the Microsoft Agent Framework. Existing container images will no longer receive updates after that date.

---

Quick start snapshot (from the README)

# scaffold a chat‑bot flow
pf flow init --flow ./my_chatbot --type chat

# add your OpenAI key
pf connection create --file ./my_chatbot/openai.yaml \
    --set api_key=<your_api_key> --name open_ai_connection

# run the bot interactively
pf flow test --flow ./my_chatbot --interactive

---

Who might use this?

• Data scientists / ML engineers building LLM‑driven products who want a structured, reproducible pipeline.
• Developers who prefer a CLI + VS Code UI for prompt engineering and want built‑in tracing and evaluation.
• Teams looking to move from prototype to production within Azure’s ecosystem (though they should plan migration to the Agent Framework before 2027).

---

License & contribution

• MIT license.
• Contributions accepted via pull requests; a CLA is required.
• Community channels: GitHub issues, discussions, and a VS Code marketplace extension.