1. Conversational Q&A

Generative Question & Answer on Private Knowledge Base using RAG(Retriever Augmented Generation)

RAG (Retriever Augmented Generation) uses a private enterprise knowledge base (like support documentation, books, contract documents, and corporate policy) to retrieve relevant parts using a neural search and use it to generate cogent & fluent output using LLM. A simple Q&A pipeline using APIs does not perform well in practice on accuracy metrics of specific Q&A and requires dedicated efforts. This repo covers the build of end to end pipeline to get generative output from multiple source documents.

Single-Turn RAG(Retriver Augmented Generation)

Ml Pipeline

Evaluation Leaderboard

• flan-t5-xxl outperforms ul2 irrespective of retriever used
• ES scores better in comparison to Solr as retriever for flan-t5-xxl
• There is a drop in accuracy when we use reranker in comparsion to when we do not.