검색 증강 생성(RAG)

Retrieval-Augmented Generation gives an LLM access to external knowledge at query time. Instead of relying only on training data, you fetch relevant text and add it to the prompt.

• Answers stay current without retraining
• Reduces hallucinations
• Lets the model cite your private documents

A typical pipeline has two phases:

• Indexing: split documents into chunks, embed them, store vectors
• Retrieval + generation: embed the query, find similar chunks, feed them to the LLM

Split long documents into smaller chunks (often 200-500 tokens) with slight overlap. Good chunking keeps related ideas together so retrieval returns coherent context.

An embedding model turns text into a numeric vector. Similar meanings produce nearby vectors. You embed every chunk during indexing.

Vectors live in a vector database such as pgvector, Pinecone, or Qdrant. Each record stores the vector plus metadata (source, title, chunk id) for later filtering and citation.

At query time you embed the user question and run a similarity search (cosine distance) to get the top-k closest chunks.

Insert the retrieved chunks into the prompt as context, then ask the model to answer using only that context.

Because each chunk carries metadata, you can show citations next to the answer. This builds trust and lets users verify claims against the original document.

If similarity scores are all low, the knowledge base probably lacks the answer. Detect this with a threshold and have the model reply that it does not know, rather than guessing.

When source documents change, re-embed and upsert the affected chunks. Track a content hash per chunk so you only re-index what actually changed, saving embedding cost.

Measure two things: retrieval quality (did we fetch the right chunks?) and answer quality (is the response grounded?). Use a test set of question and answer pairs and check whether the cited chunks contain the supporting facts.

Check your understanding of RAG.

You learned the full RAG flow:

• Chunk and embed documents into a vector store
• Embed the query and retrieve top-k similar chunks
• Augment the prompt and answer with citations
• Handle low-confidence matches and keep the index fresh

RAG grounds your AI features in your own data without retraining.