Glossary

RAG (retrieval-augmented generation)

RAG (retrieval-augmented generation) is a method where AI first searches your company's knowledge base and documents for articles relevant to the question, and only then writes an answer based on them. The model doesn't need retraining: update an article, and the AI answers from the new version.

Why it mattersAI answers according to your rules and knowledge base, not the model's guesses.

Also called: RAG, Retrieval-augmented generation, Knowledge-grounded generation, Retrieval-based generation, Source-grounded generation

How it works

Knowledge base articles are split in advance into small chunks and stored in a search index. When a customer asks a question, the system finds a few of the best-matching chunks – by meaning and by keywords – and adds them to the prompt for the language model. The model writes its answer only from those chunks and can show which article it came from.

Index: articles, policies, pricing, FAQs
Search: by meaning and by keywords
Answer: the model builds it from the chunks it found
Source: the agent sees which article the answer is based on

Use cases

Why a contact center needs it

Without RAG, a model answers from what it memorized during training and can easily make up a shipping fee or a return window. With RAG, it answers from your approved materials, and you fix an answer by fixing the article.

Without RAGThe model's general knowledge: prices, timelines, and policies may be made up or outdated.

With RAGAnswers from your knowledge base: easy to verify and update in one place.

Best practices

Short articles, one question each, with a clear title
Delete outdated articles instead of keeping them next to new ones – otherwise the AI will find both
Tell the model to say "I don't know" and bring in an agent when there's no matching article
Test the AI on real customer questions and fix the article behind every wrong answer

In CONTACT.WORK

How it works in CONTACT.WORK

AI agents and the AI assistant answer from the platform's single knowledge base, the same one agents use. If an article has an error or a gap, the AI will repeat it, so launching an AI agent starts with the knowledge base.

If the platform runs on your servers, the knowledge base and the language model stay with you too. The platform itself is free. The language model for these answers is billed separately, or you run it on your own GPU.