README.md

Question Answering Application for Healthcare

This is a streamlit-based NLP application powering a question answering demo on healthcare data. It's easy to change and extend and can be used to try out Haystack's capabilities.

A video presentation of this demo is available on YouTube. To get started with Haystack please visit the README or check out our tutorials.

Usage

The easiest way to run the application is through Docker compose. From this folder, just run:

docker compose up -d

Docker will start three containers:

• elasticsearch, running an Elasticsearch instance with some data pre-loaded.
• haystack-api, running a pre-loaded Haystack pipeline behind a RESTful API.
• ui, running the streamlit application showing the UI and querying Haystack under the hood.

Once all the containers are up and running, you can open the user interface pointing your browser to http://localhost:8501.

Screencast

https://user-images.githubusercontent.com/4181769/231965471-48d581a2-e1aa-4316-b3a4-990d9c86800e.mov

Evaluation Mode

The evaluation mode leverages the feedback REST API endpoint of haystack. The user has the options "Wrong answer", "Wrong answer and wrong passage" and "Wrong answer and wrong passage" to give feedback.

In order to use the UI in evaluation mode, you need an ElasticSearch instance with pre-indexed files and the Haystack REST API. You can set the environment up via docker images. For ElasticSearch, you can check out our documentation and for setting up the REST API this [link](https://github.com/deepset-ai/haystack/blob/main/README. md#7-rest-api).

To enter the evaluation mode, select the checkbox "Evaluation mode" in the sidebar. The UI will load the predefined questions from the file [eval_labels_examples](https://raw.githubusercontent.com/ deepset-ai/haystack/main/ui/ui/eval_labels_example.csv). The file needs to be prefilled with your data. This way, the user will get a random question from the set and can give his feedback with the buttons below the questions. To load a new question, click the button "Get random question".

The file just needs to have two columns separated by semicolon. You can add more columns but the UI will ignore them. Every line represents a questions answer pair. The columns with the questions needs to be named “Question Text” and the answer column “Answer” so that they can be loaded correctly. Currently, the easiest way to create the file is manually by adding question answer pairs.

The feedback can be exported with the API endpoint export-doc-qa-feedback. To learn more about finetuning a model with user feedback, please check out our [docs](https://haystack.deepset.ai/usage/ domain-adaptation#user-feedback).

Query different data

If you want to use this application to query a different corpus, the easiest way is to build the Elasticsearch image, load your own text data and then use the same Compose file to run all the three containers needed. This will require Docker to be properly installed on your machine.

Running your custom build

Once done, modify the elasticsearch section in the docker-compose.yml file, changing this line:

 image: "julianrisch/elasticsearch-healthcare"

to:

 image: "my-docker-acct/elasticsearch-custom"

Finally, run the compose file as usual:

docker-compose up

Development

If you want to change the streamlit application, you need to setup your Python environment first. From a virtual environment, run:

pip install -e .

The app requires the Haystack RESTful API to be ready and accepting connections at http://localhost:8000, you can use Docker compose to start only the required containers:

docker-compose up elasticsearch haystack-api

At this point you should be able to make changes and run the streamlit application with:

streamlit run ui/webapp.py

Using GPUs with Docker

Assuming you have nvidia drivers installed on your machine, you can configure docker to use the GPU for the Haystack API container to speed it up. First, configure the nvidia repository as described here: https://nvidia.github.io/nvidia-container-runtime/. For example:

curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | \
  sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | \
  sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list
sudo apt-get update

Then, install nvidia-container-runtime as described here: https://docs.docker.com/config/containers/resource_constraints/#access-an-nvidia-gpu. For example:

sudo apt-get install nvidia-container-runtime

Restart the Docker daemon (or simply the machine). Finally, you can change the docker compose file healthcare/docker-compose.yml so that a docker image prepared for usage with GPUs is used and one GPU is reserved for the Haystack API container:

  haystack-api:
    image: "deepset/haystack:gpu-v1.14.0"
    ports:
      - 8000:8000
    restart: on-failure
    volumes:
      - ./haystack-api:/home/node/app
    environment:
      - DOCUMENTSTORE_PARAMS_HOST=elasticsearch
      - PIPELINE_YAML_PATH=/home/node/app/pipelines_biobert.haystack-pipeline.yml
    depends_on:
      elasticsearch:
        condition: service_healthy
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]