How does grounding work?
Google’s machine learning glossary defines grounding as basing all or part of an LLM’s response on information retrieved from one or more trusted sources, and names retrieval-augmented generation (RAG) as a common technique. The system finds relevant information, adds it to the prompt and tells the model to answer from it.
Sources can be a company’s documents, a database, a search engine or a map service. Google Cloud’s grounding documentation lists three benefits: fewer hallucinations, responses anchored to your data, and links back to the sources so answers can be audited.
Grounded doesn’t mean correct. Google’s glossary notes that a grounded model isn’t always a factual one, because its source material, such as a physics textbook, could contain mistakes. An agent grounded on an outdated help article can repeat what the article says, so the sources need checking as much as the model does.
Why does grounding matter for a voice agent on a website?
Customers ask about your product, which a general model knows little or nothing about, or knows an older version of. Without grounding, the agent fills the gap with plausible guesses. With it, answers follow the sources you provide, so keeping those sources current matters as much as choosing the model.
Who controls the sources matters too. If anything published to a docs folder reaches the agent automatically, a draft, a mistake or a planted instruction can reach customers the same way.
How does Kawl Now ground its answers?
In Kawl Now, the sources are the knowledge your team approves: FAQs and notes you write, plus selected .md, .mdx and .txt files from GitHub for the support agent. The sales agent uses only approved sales notes. Adding a source never approves it, and every GitHub re-sync needs approval again.
During a call, the agent is told to search the approved knowledge for its channel before answering product questions, and to use only supported facts. It doesn’t search the web, read your live website or see customer accounts, and website URL import is switched off.
Two limits are worth knowing. Callers don’t see which source an answer came from, and neither does the dashboard, so you check grounding by reading transcripts against your sources. And these are instructions to the model, so they guide its answers rather than guarantee them.
Common questions
What is the difference between grounding and RAG?
Grounding is the goal: answers based on trusted sources. Retrieval-augmented generation is one way to reach it, by searching a knowledge base and adding the results to the prompt. Grounding can also draw on other sources, such as a search engine or a database query.
Does grounding stop an AI from making things up?
It makes it less likely. A grounded model can still add details its sources don’t contain, or repeat a source that is wrong, so test with real questions and review the answers.