Course contentsShow
← AI Engineering
Lesson 326 of 1,8869. Retrieval-Augmented Generation (RAG) - BasicsFree lesson

The Three-Step RAG Pipeline

Breaking down RAG into retrieval, augmentation, and generation stages with clear data flow between components.

The Three-Step RAG Pipeline

What you'll learn: How RAG breaks down into three distinct stages that work together to generate informed, context-aware responses.

Now that you understand what Retrieval-Augmented Generation is, let's examine how data flows through the pipeline. RAG isn't a single magical operation—it's a three-stage process where each component has a clear job.

Step 1: Retrieval

First, your system takes the user's question and searches your vector database for the most relevant documents or chunks. Using the embedding and search techniques you've learned, you retrieve the top-K most similar pieces of information. Think of this as a researcher pulling the most relevant books from a library before answering a question.

Input: User query
Output: Ranked list of relevant document chunks with similarity scores

Step 2: Augmentation

Next, you take those retrieved chunks and combine them with the original user query to create an enriched prompt. This is where you're "augmenting" the question with context. You're essentially telling the language model: "Here's the question, AND here's the specific information you should use to answer it."

Input: Original query + retrieved documents
Output: Enhanced prompt with context

Step 3: Generation

Finally, you send the augmented prompt to your language model (like GPT or Claude). The model generates a response based on both its training and—critically—the specific context you provided. The retrieved information grounds the answer in your actual data rather than just the model's general knowledge.

Input: Augmented prompt
Output: Final generated answer

The Data Flow

Query → Vector Search → Relevant Docs → Combine with Query → LLM → Final Answer

Each stage depends on the previous one working correctly. Poor retrieval means irrelevant context. Poor augmentation means the model won't use the context effectively. Poor generation means the answer won't satisfy the user.

Key Takeaway: RAG is a three-step pipeline—retrieve relevant information, augment the query with that context, then generate an informed response—where each stage transforms the data for the next.