RAG (Retrieval-Augmented Generation) means: when someone asks a question, your system searches your documents for relevant chunks, then feeds those chunks to an LLM to generate an answer. It's like giving the AI an open-book exam. Read the LangChain docs to understand the concept.

You need somewhere to store your documents as embeddings (numerical representations). Pinecone has a free tier that works for small projects. Alternatively, Chroma runs locally for free. This is where most beginners get stuck — embedding models, dimensions, and indexing strategies matter.

Split your documents into chunks (usually 500-1000 tokens). This is where the art comes in — chunk too small and you lose context, too large and retrieval gets noisy. Use OpenAI's embedding API to convert chunks into vectors and store them in your database.

Wire it together: user question comes in, get its embedding, search your vector DB for the top 5 most similar chunks, pass those chunks plus the question to an LLM, return the answer. LangChain has templates for this, but you'll spend most of your time tuning retrieval quality.