The most important technology behind an AI agent that provides answers that are correct.
RAG stands for retrieval augmented generation. For each question, the language model is first provided with the relevant pieces from your own documents and formulates the answer from that, instead of from general training knowledge. This way the answer remains current, verifiable and traceable to a source.
IN COMMON LANGUAGE
Look up, then answer
Without RAG, a model answers from what it once saw during training. It knows nothing about your prices, conditions or delivery times, so it starts gambling.
With RAG, something else happens first: the question is used to look up the relevant passages in your documents. These passages serve as context, and only then does the model formulate the answer.
The difference is big. If you change a price in your source material, tomorrow's answer will be correct. Without RAG you would have to retrain the model, which is slow and expensive.
WHY IT MATTERS
Why this is the standard
For almost any business application, this is the right setup: you want answers about your business, not about the world in general.
It also makes it verifiable what is happening. If an answer is incorrect, the cause can be traced back: the source was outdated, or the correct document was not found. That can be repaired, unlike a model that made something up.
- Answers from your own documents — prices, conditions, manuals.
- Immediately up to date — you adjust the source, not the model.
- Show source — so that an answer is verifiable.
- "I don't know" is allowed — better than a plausible fiction.
- Quality depends on your documentation — messy sources give messy answers.
FREQUENTLY ASKED QUESTIONS
More about RAG
Does my documentation need to be in order?
Yes, and that is the fair threshold. An agent can only answer what is written somewhere. Conflicting or outdated documents yield conflicting answers. In practice, cleaning up source material is often the real work of an AI project.
Do my documents remain mine?
Yes. They are stored in an environment that is yours and can be consulted per question. They are not used to train the model, and upon termination everything remains with you.
What if the right document is not found?
Then the officer should say that he doesn't know and refer to a human. We consciously adopt this behavior: an honest "I will find out for you" is better than an answer that sounds plausible and is wrong.
CONTINUE READING
Of course, continue reading about this subject.
Answers out your sources?
Tell us which questions are asked and where the answers are. We charge what an agent delivers.