Course Code: ragbspk
Duration: 21 hours
Prerequisites:

Audience

  • IT professionals 
Overview:

By the end of this training, participants will be able to:

  • grasp programming skills of implementing Retrieval-Augmented Generation (RAG) workflow for private knowledge base queries.

 

Course Outline:
A. Background of Knowledge Base Queries
  1. Overview of textual data search for private knowledge base queries, e.g. RAG architecture, vector database and Large Language Model (LLM)
  2. Key components, their relationships and process

B. Setup and Configuration of LLM
  1. Overview of different On-Prem LLM models
  2. Installation of Python, Hugging Face, LlamaIndex, Mistral Large 2 or similar that works with LlamaIndex, and essential libraries on local machine

C. Setup of Knowledge Base
  1. Introduction of different types of document loaders and different embedding models
  2. Programming on loading documents
  3. Programming on chunking documents
  4. Programming on transforming into embeddings, using BGE-EN-ICL or similar that works with LlamaIndex

D. Setup and Configuration of Vector Database (VDB)
  1. Overview of different VDB brands for Microsoft Server environment
  2. Installation of Postgre (pgvector) or similar that can be run on MS OS environment
  3. Programming on loading vector embeddings to VDB
  4. Programming on populating and updating VDB for new documents

E. RAG Workflow
  1. Programming on setting up query
  2. Programming on turning query into embeddings
  3. Programming on retrieving relevant chunks
  4. Programming on creating prompt with chunks and passing it to LLM
  5. Programming on generating response in LLM

F. Testing and Optimising
  1. Testing RAG workflow
  2. Optimising retrieval and generation performance

G. Hardware Requirements Covering Systems Tracks
  1. Overview of hardware requirements for the whole setup

Models

(a) LLM: Mistral Large 2 or similar that works with LlamaIndex

(b) VDB: Postgre (pgvector) or similar that can be installed under MS OS environment

(c) Embedding model: BGE-EN-ICL or similar that works with LlamaIndex

(d) LlamaIndex