Skip to content

Latest commit

 

History

History
87 lines (55 loc) · 4.28 KB

README.md

File metadata and controls

87 lines (55 loc) · 4.28 KB
Raw

Introduction

ReliefWeb Promptflow Demo

This repo contains sample code for using Promptflow as part of operationlizing LLMs for humanitarian response, as referenced in Blog post A Humanitarian Crises Situation Report AI Assistant: Exploring LLMOps with Prompt Flow

The example flow does the following ...

  1. Extracts entities from user input and converts them to a query on ReliefWeb
  2. Runs query again Reliefweb API to get situation reports for the user's request
  3. Summarized the response
  4. Answers the user question
  5. Extracts references
  6. Presents results to the user

It is a very basic app and hasn't been tuned for production use, more work would be needed for working with the ReliefWeb API. It is meant to demonstrate various things to consider when operationalizing LLM solutions.

The flow also includes:

  1. Content safety filtering
  2. Prompt variants
  3. Groundedness checks
  4. Dynamic grounding using deepeval

And the repo includes GitHub actions to run ...

  1. Promptflow automated groundedness tests
  2. Code quality tests

Setup

For setup with screenshots, see also the blog post.

Environment

  1. Install miniconda by selecting the installer that fits your OS version. Once it is installed you may have to restart your terminal (closing your terminal and opening again)
  2. In this directory, open terminal
  3. conda env create -f environment.yml
  4. conda activate promtpflow-serve

The repo should support both OpenAI and Azurer OpenAI depending on the variables set in the .env file. If you want to test content safety, you will need to set up an Azure content safety instance, or dicoennect that connection in the flow and implement your own custom solution.

Promptflow

Promptflow can be run from the commandline, see documentation for further information, but a nice way to use it is to use VS Code which has a user interface for managing flows. To use this ..

  1. Download VS Code
  2. Install the promptflow extension
  3. Install the conda environment (see above)
  4. Open a flow.dag.yaml
  5. At top of file, click install dependencies
  6. Select the conda environment promptflow-serve
  7. Re-open flow.dag.yaml, select 'visual editor' at the top to see the lovely user interface
  8. To run the flows click the play icon at the top of the promptflow user interface

You will also need to configure LLM keys. The demo assumes Azure OpenAI, but scripts can also support OpenAI direct. To configure your LLM environment ...

  1. Copy .env.example to .env
  2. Set keys appropriately

Azure versus OpenAI

The code is configured to run with Azure OpenAI. You can also run with OpenAI directly as follows:

  1. In promptflow, create a new OpenAI conntection (in VS code select 'P' promptflow on left,click + under connections). For command line creation, see .github/test_deploy,yml)

  2. Set connection in all LLM nodes in the flow using VS code (click on them, change connection)

  3. In deep_eval.py adjust code to use OpenAIChat instead of AzureOpenAI. At some point this will be a settings

Note, if using promptflow in Azure ML, you can explore other model connections. After creating an appropriate deployment, follow the steps above.

Development

Pre-commit hooks

The repo has been set up with black and flake8 pre-commit hooks. These can be configured in the .pre-commit-config.yaml file and initialized with pre-commit autoupdate.

On a new repo, you must run pre-commit install to add pre-commit hooks.

To run code quality tests, you can run pre-commit run --all-files

Automatic testing using Promptflow

Automatic tests are run using Github actions, which creates a promptflow connection and executes a promptflow evaluation run. The output is monitored by a script, and can be used as a template for adding promptflow tests as part of DevOps.

See ./github/workflows/test_deploy.yml for more details, and 'Actions' in the repo