Mono QA

This repository contains the code to run the mono-qa server.

Prerequisites

Python

I need python >= 3.6 for everything to work here, as well as pip and venv.

We also require an instance of ElasticSearch to be running. It assumes that it will be running on http://localhost:9200. You can download ElasticSearch here

Currently I'm running on ElasticSearch 7.7

pyTorch

pyTorch is used here. On linux, this can be installed with pip no problem. On windows, it's a bit more complicated but the fellas at pytorch will provide you an installation command here.

BE SURE TO ACTIVATE YOUR VIRTUAL ENVIRONMENT FIRST!

At this very moment, the installation command for windows is:

pip install torch===1.5.0 torchvision===0.6.0 -f https://download.pytorch.org/whl/torch_stable.html

Usage

You can go to the src directory and run the config script:

# Linux/bash

cd src
./config.sh

# Windows/powershell

sl src
powershell config.ps1

This will set up a virtual environment and install required dependencies. To use the program you will need to activate your virtual environment:

# Linux/bash

source bin/activate

# Windows/powershell

powershell Scripts/Activate.ps1

Now you can run the server:

python server.py

API

As of now, the server will accept POST requests at

http://192.168.0.72:8000/question

The request should look like:

POST /question HTTP/1.1
host: 192.168.0.72:8000
content-type: application/json

{
    "question": "where is your team based?"
}

Or, in cURL:

curl http://192.168.0.72:8000/question -H 'content-type: application/json' -d '
{
    "question": "where is your team based?"
}
'

You will then receive a json response that looks something like:

{
    "quick_answer": "We are located in southern Europe.",
    "question": { 
        "text": "where is your team based?",
        "uuid": "24f8619a-df94-4104-869f-22ecf0221ede"
    },
    "answers": [
        { "answer": "We are located in southern Europe.",
          "rating": .95,
          "paragraph": "Mono is a company...",
          "paragraph_rank": 1,
          "docId": "wltzwnIBO-Ft5Ide2Lj7"
        },
        ...
    ]
}

The quick_answer is provided for convenience and will correspond to the highest rated answer in the answers array. It is possible that no answer is found, in which case you will get a json that looks something like:

{
    "quick_answer": "",
    "question": { 
        "text": "where is your team based?",
        "uuid": "24f8619a-df94-4104-869f-22ecf0221ede"
    },
    "answers": [
        { "answer": "",
          "rating": 0.8,
          "paragraph": "Mono is a company...",
          "paragraph_rank": 1,
          "docId": "wltzwnIBO-Ft5Ide2Lj7"
        },
        ...
    ]
}

A few important things to note:

• quick answer is the empty string\
• the answers array is not empty\
• answer in the answers array has a non-zero rating\

The QA model may decide that a paragraph does not contain the answer to a question. This means it will give an empty span the highest rating of all spans. For feedback mechanisms we may wish a user to look at the paragraphs and decide if in fact the answer is there, or if the wrong answer is chosen (i.e. the quick_answer is wrong), but the "correct" answer corresponds to a different paragraph, this information may be valuable.

In some (probably rare cases) it may happen that ElasticSearch finds no paragraphs related to some question. In this case, the answers array will be empty. Be sure to be prepared to handle this case.

{
    "quick_answer": "",
    "question": { 
        "text": "gdje se nalazi tvoja ekipa?",
        "uuid": "24f8619a-df94-4104-869f-22ecf0221ede"
    },
    "answers": []
}

Contact

The original author of this code can be reached at:

nchappell@mono-software.com