16 docs tagged with "RAG"

100 RAG interview questions with detailed answers — the complete study guide for AI Engineer interviews. Fundamentals, chunking & embeddings, vector DBs, evaluation, and production RAG.

My notes for the Ultimate RAG Bootcamp (LangChain · LangGraph · LangSmith) — organized section-by-section exactly like the course, from RAG foundations to agentic, corrective, adaptive RAG and GraphDB. Updated as I progress.

Semantic chunking — splitting where the meaning shifts instead of at fixed sizes — how it works under the hood, the threshold knob, and building a RAG pipeline on top of it.

The building blocks every RAG system is made of — loaders, splitters, embeddings, vector store, retriever, and the LLM — and how they connect.

Loading and parsing every source type into LangChain Documents — text, PDF, Word, CSV/Excel, JSON, and SQL — plus project setup, the Document structure, and text splitting. One submodule per topic, ending with a cheat sheet.

Combining dense (vector) and sparse (BM25) retrieval with an ensemble, then sharpening results with cross-encoder reranking and diversifying them with MMR.

What the Ultimate RAG Bootcamp covers and how it's structured — from traditional RAG to advanced, multimodal, and agentic RAG with LangChain, LangGraph, and LangSmith.

What RAG is and why it exists — grounding an LLM in external data at query time to fix stale knowledge, missing private data, and hallucination.

The four Lang* tools in one place — what each is, how they differ, and when to use which. LangChain builds it, LangGraph orchestrates it, LangSmith watches it, Langflow lets you draw it.

Building workflows as graphs — the State schema, nodes as functions, edges and conditional (router) edges, compiling and invoking a StateGraph, the add_messages reducer, and a tool-calling chatbot with ToolNode + tools_condition.

Retrieval over documents that mix text and images — extract the text, have a vision model describe each image as a caption, embed both into one vector store, and answer from whichever is relevant.

Reshaping the user's query before retrieval — query expansion (synonyms/context), query decomposition (split complex questions), and HyDE (embed a hypothetical answer) — so the retriever finds better context.

The LangChain v1 essentials — create_agent, one init_chat_model call across OpenAI/Gemini/Groq, the message types, tool binding and the tool-calling loop, structured output with Pydantic/TypedDict, and middleware like summarization.

What embeddings are and how to create them — the meaning-vector intuition, cosine similarity, HuggingFace and OpenAI embedding models, embedding single queries vs batches of documents, and choosing a model.

Vector store vs vector database, then building a full traditional RAG pipeline with ChromaDB and LangChain — load, split, embed, store, similarity search, and answer with the modern chain and LCEL.

The local dev setup for the course — VS Code, Anaconda/conda environments, and keeping each project isolated with its own dependencies.