Generative AI with Context: RAG, CAG & KAG Applications

Install git on your local system to clone repositories using git bash. Follow platform-specific download steps for Windows, macOS, or Linux, and complete the setup wizard to finish the installation.

Explore the course repository with three projects on retrieval, cache, and knowledge augmented generation. Clone the repo, set up a Python virtual environment, and use VSCode to run the exercises.

Set up environmental variables by creating a dot env file and storing your OpenAI API key, optionally a secondary key for other models, to connect Python code with the backend.

Explain retrieval augmented generation, including document ingestion, embedding, vector store, semantic search, context augmentation, and grounded response to overcome base LMS limitations.

Explore how to source proprietary data and chunk documents into complete, meaningful pieces to improve context windows, retrieval accuracy, granular search, and scalability.

Convert text into numeric vectors to capture meaning. Place document chunks in an n-dimensional space to reveal semantic similarity and enable semantic search.

Explore how a vector store organizes embedding chunks by meaning and enables semantic search to retrieve the closest, contextually relevant text chunks using cosine similarity.

Retrieve the most relevant context from the vector store, perform context augmentation by merging it with the user question, and generate a grounded answer from the large language model.

Learn the Rac project structure, including the knowledge base, raw and rack files, and setup for a Python OpenAI workflow. Review embeddings, cosine similarity, semantic search, and embedding cache.

Open the rack project and initialize the OpenAI client by loading the API key from a .env file. Import os, pathlib, dotenv, and numpy to support embeddings and chat completion.

Create and load a knowledge base by scanning the knowledge directory for md and pdf documents, building a dictionary of document IDs to contents, and gracefully handling missing directories.

Define an embedding function that converts input text into a vector using OpenAI's text embedding tree small model and returns the first embedding from the response data.

Explore cosine similarity on embedding vectors to measure the similarity between a user query and text, using dot products and magnitudes, and build an embedding cache for semantic search.

Explore building an embedding cache for a knowledge base by encoding documents with a get_embedding function, caching vectors in a dictionary, and enabling fast cosine similarity-driven semantic search.

Dive into semantic search by embedding the query and comparing it to document embeddings stored in the embedding cache, then retrieve top-k most similar knowledge base entries using cosine similarity.

Develop a rag pipeline that searches relevant knowledge base documents using semantic search and cosine similarity, builds context, augments prompts, orchestrates steps, and generates answers with OpenAI's large language model.

Load the knowledge base, cache embeddings, and run rag queries with semantic search and cosine similarity to retrieve top documents and answers.

Explore the exercise structure for building an augmented generation system using loading documents, embedding models, cosine similarity, and semantic search, with rag query and a bookstore knowledge database.

Demonstrate a retrieval augmented generation exercise solution with a library chatbot, connecting OpenAI to databases of science fiction, mystery, and children's books, and returning top documents by similarity.

Explore cache augmented generation (CAG), which preloads cache knowledge with a query for fast, simple grounded answers, and compare it to retrieval-based RAG for dynamic vs static knowledge.

Explore cached augmented generation (CAG) by loading and caching the knowledge base with the query, then feeding it to a large language model, and study project two’s structure.

Load knowledge base shows how CAG combines all documents into one big string to form the full context, contrasting with RAG's dictionary separation for search.

Develop the CAG pipeline by loading the knowledge base, building the cat query function, and using system prompts with OpenAI chat completion to answer user questions from knowledge context.

Develop a CAG-based movie recommendation assistant using a small to medium knowledge base, with guided steps to load knowledge, query, and implement main functions, contrasting with RAG for larger datasets.

Demonstrate generative AI with context through a movie recommendation project that uses a knowledge base, a system prompt, and prompt engineering to deliver database-based recommendations.

Explore knowledge augmented generation (KAG) and its knowledge graphs that link concepts through relationships and reasoning for more coherent, accurate answers.

Learn how knowledge graphs map facts, entities, and edges and power knowledge augmented generation by feeding a JSON file to search and answer with an LLM.

Learn the knowledge augmented generation workflow using a knowledge graph, with entity linking, graph traversal, and knowledge extraction, then ground the LLM and compare with RAG.

Learn the knowledge augmented generation project structure, including generating graphs from documents, loading and using a knowledge graph, and applying RAG and CAG workflows.

Generate a knowledge graph from documents by extracting entities (nodes) and relationships (edges) with subject, verb, and object, encoded as json via a prompt to a large language model.

Load knowledge graph by reading the generated graph JSON and extracting entities and relationships. Prepare data to support RAG and CAG workflows.

Extract entities from the user question by matching lowercase text against knowledge graph entities. Use a name map for nicknames and variations, then append matched entity IDs to found entities.

Extract facts from the knowledge graph by gathering entity properties and relationships, formatting them into readable text, and handling subject, predicate, and object relations for augmented generation.

Fuse knowledge graph facts with RAG methods by embedding texts, computing cosine similarity, and semantic search in a CAG query pipeline.

Develop a kag pipeline that extracts entities, retrieves knowledge graph facts, performs rag-based document search, fuses results, and queries a large language model to generate answers.

Load the documents, generate or load the knowledge graph from them, and save the graph for future use, then run the cat query pipeline to answer questions.

Explore building a music-recommendation chatbot using a knowledge graph and RAG pipeline, covering document loading, graph generation, entity extraction, fact retrieval, and hybrid fusion with a large language model.

Learn to build and run a knowledge-augmented generation workflow with RAG, CAG, and KAG by generating and loading a knowledge graph from documents, then fusing contexts to answer queries.