podmind - virtual conversations with podcast guests

Here is the live system: Voxgig Podcasts.

This is an open source example of a Retrieval Augmented Generation (RAG) system that includes:

• A document (audio transcript) ingestion pipeline;
• Simple vector embeddig using AWS LLMs and OpenSearch;
• An embeddable chat widget that can run on third party websites;
• A scalable Node.js Serverless Microservices architecture in TypeScript;
• An exmaple of the local-monolith development style;
• An example of a type-safe model-driven system;

This code is derived from code written by Mikael Vesavuori: bedrock-rag-demo (Thanks Mikael!)

	This open source code is sponsored and supported by Voxgig.

Requirements

Locally the system runs as a single Node.js process (a "local monolith").

You will need a Deepgram account to perform audio transcriptions. Register for a free developer account.

You will also need an AWS account to access the Bedrock LLM.

Running Locally

Clone this repo, and run npm install in these folders:

• backend
• widget/ask

Add you Deepgram API Key to the srv/env/local/local-env.ts file and recompile with npm run build

Set your $AWS_PROFILE environment variable so that the AWS SDK can access the Bedrock LLM.

In backend run:

$ npm run local

Verify the backend by opening another terminal in the backend folder and running the system REPL:

$ npm run repl

Using the REPL, ingest a podcast:

> aim:ingest,subscribe:podcast,feed:'https://feeds.resonaterecordings.com/voxgig-fireside-podcast'

This is a Seneca microservices system.

To chat to the system, run the chat widget in development mode:

In the widget/ask folder:

$ npm run dev 

And open a browser as directed.

Deployment

This system uses the Serverless framework for deployment to AWS.

The following AWS services are used:

• Lambda
• DynamoDB
• OpenSearch
• S3
• Cloudwatch
• SQS
• API Gateway
• Cloudfront

Basic AWS and Serverless knowledge is assumed.