Intro to Algorithms Class Chatbot

Prerequisites for Development

python 3.9 and pip
- tensorflow
- tensorflow_hub
- tensorflow-text
- (See details in "How to Run the Project" about installing other dependencies)
Haystack (see installation instructions)
- Apple M1 has extra dependencies
Elasticsearch host (see details in "Miscellaneous" about elasticsearch)

Prerequisites for Deployment

Docker
Elasticsearch host (see details in "Miscellaneous" about elasticsearch)

How to Run the Project for Development

Change directory to src/server
Run shell script run.sh to prepare and store in your elasticsearch host the information of
- i) question and answer (Q&A) pairs,
- ii) algorithm general knowledge base and
- iii) syllabus knowledge base.
Install pipenv by pip install pipenv
Run pipenv install to install backend dependencies listed in the Pipfile.lock
Run python -m src.api.app to start the Flask API, which then initiates the NLP pipeline

Project Deployment

Change directory to src/server
Run shell script prod.sh to build an image of the API and the NLP pipeline. Then, the backend will be running in a docker container. It has been tested that in Linux x86 system the docker image can be successfully built.
It is suggested that backend server should have a GPU for faster inferencing time.

Architecture

NLP pipeline visualization:

Performance of QueryClassifier

Precision = 1.00
Recall = 0.83
F1 = 0.91
Confusion Matrix:

Miscellaneous

In Fall 2022, the backend is hosted on the Google Cloud Platform. We utilized a Nvidia Tesla P100 (225W, 16G RAM) and 2 vCPU with 7.5G RAM. This results in 5s response time for each request on average.
- A elasticsearch host should be running locally or on the cloud. We used to have a free-trial cloud elasticsearch host provided by elastic.co. Alternatively, elasticsearch instance can also be hosted in a docker container.