Retrieval Augmented Generation

#digitalgarden #molecule ##emergingtechnology

RAG

Retrieval Augmented Generation, or "RAG", is an informatic tool that enables a Large Language Model (LLM) to "retrieve" information from a Knowledge Base.

In other words, you can use Natural Language Processing (normal words) to ask questions of the AI (LLM) and the LLM will pull information from a knowledge base to answer questions.

Library Analogy

Using RAG is like going to the library with a question. You bring your question to the librarian (the LLM) who goes and pulls the best books and identifies the best chapters/quotes to answer your question. The librarian then compiles the information in a way that best answers your question, "generating" a response.

The library in this case can be any Knowledge Base; articles, youtube videos, personal notes, company data, etc.

Key Elements Of RAG

Chunking: splitting the text into segments, standalone units that retain context

Embeddings: organizing the chunks based on their relevance to each other, aka based on similarity

Vector Database: stores the embeddings, creating a database that can be searched based on context, not just exact matches

Large Language Model (LLM):

  1. LLM breaks up your question into chunks to compare it against the existing database
  2. Database sends answer chunks to LLM based on similarity to what it has in its system
  3. LLM is the synthesizer that pieces the retrieved chunks together in a way that makes sense to answer the question