Science Chatbot

Architecture

1. Using Science - Wikipedia.pdf as data source.
2. Dividing pdf into chunks.
3. Created VectorDB.
4. Embed query, on top of that apply cosine similarity followed by cross encoder reranking.
5. You will get Most relevant documents.
6. Apply falcon-40b model to create an answer to query.
7. I am using FASTAPI here.

Input Schema for /chatbot - POST request

{
    "query" : "What is Science ?"
}

Output Schema

Science is ....

Requirement

The current project has the blueprint structure of an AI App.

Your mission is to implement an 💬NLP chatbot answering questions about science.

You will add your logic to the main.py file inside the execute function.

def execute(request: SimpleText, ray: OpenfabricExecutionRay) -> SimpleText:
    output = []
    for text in request.text:        
        response = '' # <<< --Your magic goes here
        output.append(response)

    return SimpleText(dict(text=output))

Constraints and restrictions

You are free to use any package or library you see feet as long as you follow these rules:

• 👎 You can't call any external service (e.g. chatGPT)
• 👎 You can't copy and paste from other peoples work

Run

The application can be executed in two different ways:

• locally by running the start.sh
• locally by running on FASTAPI uvicorn app:app --proxy-headers --host 0.0.0.0 --port 80 --reload
• on in a docker container using Dockerfile

Docker commands

docker build -t open-fabric .
docker run --gpus all --publish 80:80 -it open-fabric

Minimum requirements

1. Minimum GPU VRAM : 24 GB.

Submission

Your solution must be uploaded on GitHub, and submit us the link in max 1 week after receiving the task.

Note

Keep in mind that this is the project that will be used to evaluate your skills. So we do expect you to make sure that the app is fully functional and doesn't have any obvious missing pieces.