README.md

Building a retrival-based Question Answering system using LangChain + Ray in 20 minutes

This is source code for "Building a retrival-based Question Answering system using LangChain + Ray in 20 minutes"

Requirements

This demo requires a bit of a hefty setup. It requires one machine with a 24GB GPU (eg. an AWS g5.xlarge) or a machine with 2 GPUs (minimum 16GB each) or a Ray cluster with at least 2 GPUs available.

It probably will be too slow if you run it on your own M2 mac. Torch drivers are not well optimized for the M2 yet.

Getting started

Once you have a cloud machine with the required specifications, do pip install ray[default]

To apply the indexing to the Ray docs, use tar xvfz docs.ray.io.tar.gz

Installing dependencies

To install required dependencies, do pip install -r requirements.txt

Building the vector store index

To build the index do python build_vector_store.py

Serving

To serve do serve run serve:deployment

This sets up a service on port 8000 by default.

Querying

To query, we've included a simple python script.

python query.py 'What is the difference between SERVE and PACK placement groups?'