search

aws-samples / serverless-pdf-chat

LLM-powered document chat using Amazon Bedrock and AWS Serverless
https://aws.amazon.com/blogs/compute/building-a-serverless-document-chat-with-aws-lambda-and-amazon-bedrock/
MIT No Attribution
228 stars 206 forks source link

Amazon Kendra in place of Embedding Model #28

Closed mohsiniqbal368 closed 9 months ago

mohsiniqbal368 commented 9 months ago

Hi,

What could be the possible benefit and drawback of using Amazon Kendra for RAGing instead of using Embedding Model (Titan-1)/FAISS for similarity search. Cost could be one factor I suppose, what about the overall chat performance and answers accuracy in both cases?

pbv0 commented 9 months ago

Kendra is a fully-managed search service, so it will take care of much of the work you have to do when manually generating embeddings etc.

Regarding performance, I would suggest to try out different approaches for your specific type of documents, it's not possible to make a general assessment here.

We offer several sample for RAG with Kendra, for example here. https://github.com/aws-samples/aws-bedrock-with-rag-and-react