องค์ประกอบหลักของระบบ RAG

Let us break down a RAG system’s parts. Think of RAG as a smart librarian for your LLM, fetching the right facts from your own knowledge base.

It starts with document loading: pulling your knowledge in from any source — PDFs, web pages, databases, plain text — so the LLM can draw on it.

Document loading is like scanning your books and notes in for the AI to read. Sources span PDFs, web pages, and structured databases alike.

Loaded docs are usually too big, so you split them into chunks. Smaller pieces fit the LLM’s limited context window and are faster to search.

Each chunk becomes an embedding — a numeric fingerprint of its meaning. Similar meanings get similar fingerprints, so machines can compare ideas, not just keywords.

Those embeddings live in a vector store: a database tuned to hold and search high-dimensional vectors fast, finding the most similar chunks for any query.

When a question arrives, retrieval kicks in: embed the query, compare it to the vector store, and pull the top chunks closest in meaning.

Next is augmentation: the retrieved chunks get attached to the user’s question, handing the LLM a factual cheat sheet drawn straight from your data.

Finally, generation: the LLM reads the query plus retrieved context and writes a grounded, factual answer — sharply cutting hallucinations.

The full RAG flow: load, split, embed, and store your docs; then embed the query, retrieve top chunks, augment the prompt, and generate the answer.

Which RAG component is responsible for converting text into numerical representations that capture its meaning?

You now know RAG’s pieces: loading, splitting, embeddings, vector stores, retrieval, augmentation, and generation — working together for informed AI answers.