🛠️ WIP 🛠️

This is currently a Work in Progress, it is basically an implementation plan.

To see a working version that uses OpenAI and open source, see chat-with-confluence

Overview

The intention of this repo is to use Embedded Vector Store and an LLM to chat with your confluence docs while assuming that you only have an AWS license - no OpenAI, no other vector store service, nothing else.

Details

Approach (assuming you don't have access to Amazon Bedrock)

1. Create an embedding vector store of the Confluence docs

• use langchain.document_loaders.ConfluenceLoader to load the Confluence docs
• use langchain.text_splitter.CharacterTextSplitter and langchain.text_splitter.TokenTextSplitter to break up the docs into documents for processing
• use HuggingFace sentance transformers embeddings
• use postgres/pg_vector in RDS as the vector store

2. use the embedding vector store to get relevant texts from the store

• Create an embedding of the question (also using HuggingFace)
• Perform similarity search in the vector store using the question embedding to get relevant texts from the store

3. provide the llm with the received texts + question as a prompt to answer the question

• Use llama-2, which is available in AWS Sagemaker
• use langchain.chains.RetrievalQA

Alternative Approach assuming you have access to Bedrock 🪨

Amazon Bedrock is under "Limited Beta Preview" and it's super hard to get into the beta. I suspect this soution would be better. Here is how I would do it. The main differences are bolded

1. Create an embedding vector store of the Confluence docs

• use langchain.document_loaders.ConfluenceLoader to load the Confluence docs
• use langchain.text_splitter.CharacterTextSplitter and langchain.text_splitter.TokenTextSplitter to break up the docs into documents for processing
• (different) use langchain.embedddings.BedrockEmbeddings
• (different) use AWS OpenSearch (which is basically ElasticSearch) via langchain.vectorstores.OpenSearchVectorSearch as a vector store, following these resources:
  • langchain's docs on how to use it with Amazon OpenSearch Service Serverles
  • this announcement from AWS, which also states that more info on how to use it via langchain is forthcoming.
  • the recommendations of this guy

2. use the embedding vector store to get relevant texts from the store

• (different) Create an embedding of the question (also using langchain.embedddings.BedrockEmbeddings)
• Perform similarity search in the vector store using the question embedding to get relevant texts from the store

3. provide the llm with the received texts + question as a prompt to answer the question

• (different) use Claude 2 (recetly made available) on langchain.llms.Bedrock
• use langchain.chains.RetrievalQA

License

MIT License