What Is Naive RAG

Where we left off

If you've read andreinita.co/learning/rag-fundamentals (or this repo's other courses' RAG lessons), you've already met the idea of retrieval-augmented generation in prose: an AI model only knows what it was trained on, so RAG hands it fresh, relevant text at the moment it's asked a question, instead of retraining it. This course builds that idea, literally, starting from nothing.

Naive (Standard) RAG is the baseline version of that idea: chunk a document into passages, embed them, retrieve the ones closest in meaning to the question, and stuff them into the prompt. Every other RAG architecture in the wider taxonomy (Hybrid, Graph, Corrective, Agentic, and so on) is a variation or extension of this exact four-stage shape, so this course is the foundation the rest of the series builds on.

Why build it by hand, with no framework?

This repo's langchain, llamaindex, and pgvector courses already teach RAG through a framework or database, where a lot of the four stages above happen inside one method call. That's the right way to build a real system, but it can leave the actual mechanics a little opaque: what is embed_documents really doing? What does "similarity" mean as a number?

This course answers those questions directly: no LangChain, no LlamaIndex, just google-genai (Gemini's own official Python SDK) and a plain Python list standing in for a vector database. By the time you reach this course's Advanced tier and reintroduce a real vector library (chromadb), you'll know exactly what it's doing for you, because you'll have already done it yourself.

The code, piece by piece

from google import genai
client = genai.Client()

google-genai is Google's own SDK for Gemini, the same one LangChain's ChatGoogleGenerativeAI and LlamaIndex's GoogleGenAI classes use underneath. genai.Client() reads GOOGLE_API_KEY from your environment automatically (that's why load_dotenv() runs first) and gives you one object, client, through which every Gemini call in this course goes.

response = client.models.generate_content(model=CHAT_MODEL, contents=QUESTION)

generate_content is the raw-SDK equivalent of .invoke(): send some text, block until Gemini replies. response.text (used below) pulls out just the written answer, same idea as LangChain's AIMessage.text.

The question this lesson asks is answered inside this course's own fixtures/notes/weather-station.md file, a file Gemini has never seen. Run it and Gemini will either admit it doesn't know, or, as often happens, invent a specific-sounding but wrong answer. Both outcomes are the same underlying problem: the model can only answer from what it was trained on, and a small personal note about a Raspberry Pi weather station was never part of that training data.