Retrieval-Augmented Generation
What is it?
RAG is a technique that gives a language model relevant documents to read at question time, so its answers are grounded in real, up-to-date information.
Explain like I'm 5
Why was it created?
LLMs can be out of date and may invent facts. RAG was created to ground answers in your own or current data without retraining the model.
Where is it used?
- Chatbots over company docs
- Customer support assistants
- Search with natural-language answers
- Question-answering over private data
Why should developers care?
RAG is the most common way teams build accurate AI assistants over their own documents, so it's a key pattern for AI features.
How does it work?
Your documents are split into chunks and stored as embeddings in a vector database. At question time, the system retrieves the most relevant chunks and includes them in the prompt so the model answers using that context.
Real-world example
A support bot answers a policy question by first retrieving the exact relevant section from the company handbook, then summarizing it for the user.
Common use cases
- Answering questions over private documents
- Keeping answers current without retraining
- Reducing hallucinations
- Citing sources in responses
Advantages
- Grounds answers in real data
- No model retraining needed
- Easy to update by changing documents
- Can cite sources
Disadvantages
- Quality depends on retrieval
- Adds infrastructure (vector store, pipeline)
- Irrelevant retrieval can mislead the model
- Latency from the extra lookup
When should you use it?
When you need an LLM to answer accurately over specific or changing information.
When should you avoid it?
For general tasks where the model's built-in knowledge already suffices.
Alternatives
Related terms
Interview questions
Beginner
- What does RAG stand for?
- Why give a model documents at question time?
Intermediate
- What role does a vector database play in RAG?
- How does RAG reduce hallucinations?
Senior
- How would you improve retrieval quality?
- When would you choose RAG over fine-tuning?
Common misconceptions
- "RAG retrains the model" — it doesn't; it just supplies relevant context at query time.
- "RAG eliminates hallucinations" — it reduces them, but poor retrieval can still mislead the model.
Fun facts
- RAG stands for Retrieval-Augmented Generation.
- It relies on embeddings to find text by meaning rather than exact keywords.
Timeline
- 2020 — Retrieval-augmented generation introduced in research
Learning resources
Quick summary
RAG retrieves relevant documents and feeds them to an LLM at question time, grounding answers in real, current data without retraining.
Cheat sheet
- Retrieve relevant docs, then generate
- Grounds answers in your data
- Uses embeddings + a vector database
- Reduces but doesn't erase hallucinations