Fundamentos da Geração Aumentada por Recuperação (RAG)

An LLM only knows what it was trained on. It cannot answer questions about your private docs or recent events. RAG (Retrieval-Augmented Generation) closes this gap by fetching relevant text and putting it in the prompt.

Instead of fine-tuning the model on your data, you retrieve the most relevant snippets at query time and let the model answer using them as context. Cheaper, faster to update, and easy to cite.

Documents are split into small chunks (a few hundred tokens each). Chunks small enough to be precise, large enough to keep meaning.

Each chunk is converted to a vector with an embedding model. Similar meanings produce nearby vectors, enabling semantic search.

Vectors are saved in a vector database like Pinecone, Weaviate, or pgvector. It supports fast nearest-neighbor search over millions of chunks.

When a user asks a question, you embed the question and find the top-k most similar chunks.

The retrieved chunks are inserted into the prompt as context, and the model is told to answer using only that context.

Because the answer is built from real chunks, you can show citations back to the source documents, and you can instruct the model to say I do not know when the context lacks the answer.

• RAG updates instantly: change a doc, re-index, done.
• Fine-tuning is slow and bakes knowledge in.
• RAG gives traceable sources; fine-tuning does not.

Poor chunking, weak embeddings, or retrieving too few chunks all hurt quality. If answers are wrong, inspect what was retrieved first; the issue is usually retrieval, not the model.

• Add overlap between chunks.
• Use hybrid keyword + vector search.
• Re-rank results with a cross-encoder.
• Tune top-k for your context window.

RAG chunks documents, embeds them into a vector store, retrieves the most relevant chunks for each query, and augments the prompt. It grounds answers, enables citations, and stays current without retraining.