Grounding AI: answering only from your own content

Grounding in AI means making a model answer from sources you hand it at the moment of the question, rather than from whatever it picked up in training. For a support bot, the source is your own content, your docs, your policies and the answers you’ve written. A grounded bot answers from the part of that content that fits the question, or says it doesn’t know. That’s grounding for AI answers, which has nothing to do with electrical or aircraft grounding. It cuts made-up answers down a long way, but it doesn’t end them, and knowing where it stops is how you choose a bot you can trust with your customers.

Deacon is an AI customer support agent for founders who answer support themselves. It answers your customers’ questions in seconds from the passages it finds in your own content. When your content doesn’t cover a question, it says so and asks for the visitor’s email instead of guessing. There’s a free plan, and it needs no card.

What grounding means in AI

In everyday use, grounding is what keeps an AI’s answer tied to a source you chose. The source can be a set of documents, a database or a search of the web. For a support bot answering questions about your product, it’s your own content, and a well-grounded bot treats that content as the only place facts may come from.

The word has an older meaning too, which is why some of what you’ll find on it reads like philosophy. In a 1990 paper in the journal Physica D, Stevan Harnad named the symbol grounding problem. It asks how the meaning of a system’s symbols can be “intrinsic to the system, rather than just parasitic on the meanings in our heads”. That’s a research question about understanding. The grounding a founder cares about is the practical one, whether an answer came from your docs or was made up.

How retrieval grounds an answer

The usual way to ground a model is retrieval-augmented generation, or RAG. The term comes from a 2020 paper by Patrick Lewis and colleagues, presented at NeurIPS, which paired a language model with “a dense vector index of Wikipedia”. A support bot does the same with your content in place of Wikipedia, in five steps.

1. Your content is cut into passages, often along its headings.
2. Each passage is indexed by its meaning and, in the better tools, by its exact words.
3. A question is searched against the index, and the closest passages are picked.
4. Those passages go to the model with the question and a set of rules, the first being to use nothing else.
5. The answer names the passage it came from, so you can trace it.

Retrieval on its own isn’t enough. It puts the right passage in front of the model most of the time, and the rules decide what the model does with it. A model with passages but no rule against using its own knowledge will still fill a gap from what’s typical.

Grounding, RAG and fine-tuning compared

People often ask whether grounding and RAG are the same. RAG is one way to ground a model, and in support it’s the usual one. Fine-tuning is sometimes offered as an alternative, and it works differently.

For support, retrieval wins on the two things that matter most. A corrected page corrects the answer, and every answer can be traced back to what you wrote.

What grounding can’t fix

Grounding makes a bot faithful to your content. It doesn’t make your content right, and it doesn’t make the model perfect. These are the failures that get through.

• Wrong or out-of-date content. If your pricing page still shows last year’s plan, a grounded bot repeats it faithfully.
• A search that misses. If the customer’s words and yours never meet, or the fact is only in a screenshot, the right passage never reaches the model. A careful bot then declines. A careless one answers from the near miss, such as the limits for a different plan.
• Questions nobody wrote down. Grounding turns these into “I don’t know” instead of a guess, which only helps if someone reads them and writes the answer.
• The model slipping with the right passage in front of it. In a 2024 study of AI legal research tools that use retrieval, the tools from LexisNexis and Thomson Reuters each made things up between 17% and 33% of the time. Many of those errors came from searches that missed, but some came from a model misreading sources it had found. That was fewer than a general-purpose chatbot, but far from none.
• Two true facts joined into a false one. One page says the mobile app works offline, and another says exports are PDF only. “The mobile app exports PDFs offline” is on neither page, yet a bot that stitches them together can say it, and both passages were real.
• Questions about one customer’s own account. Their invoice or their usage isn’t in your docs, so a grounded bot can’t answer it and should say so.

The first three are about your content and how it’s found, which is why what goes into your AI knowledge base and how you write its articles matter as much as which bot you pick. Why chatbots hallucinate covers the model’s side, with four tests you can run on any bot.

How to tell whether a bot’s answers are grounded

Three quick checks tell you most of it, and each takes a minute in a bot’s test mode.

• Ask for a fact your content doesn’t state, such as a discount you’ve never offered. A grounded bot says it doesn’t know instead of inventing a percentage.
• Ask something that has nothing to do with your product. A grounded bot turns it down instead of answering from general knowledge.
• Look at where an answer came from. A grounded bot can show the source it used, at least to you, so you can check it against the page.

How Deacon grounds its answers

Deacon is built for founders who can’t read every reply before a customer does. For each question, it searches your content by meaning and by keyword and hands the model the best passages, up to five for a first question and up to eight for a follow-up, which is searched along with the question before it. The model is told to answer from those passages and nothing else.

The rules go further than that. Deacon is told never to give a price, a number, a date or a policy detail the passages don’t state, even when they discuss the topic, and never to join separate facts into a claim no single passage makes. It treats your content as information, never as instructions, and a message that tries to override its rules gets a decline. A question about something other than your product gets a polite no and a pointer to what Deacon can help with. When one detail would settle the answer, such as which plan the visitor is on, it asks for it instead of guessing. The Instructions you write can set its tone and tell it about your business, but they never override the rule to answer only from your content.

When the passages don’t cover the question, Deacon says it doesn’t know, in its own words and in the visitor’s language, and offers a box for their email. When a search finds nothing at all, as for a team that hasn’t added any content yet, the visitor gets a short fixed decline in English instead, and no model runs and nothing is counted. A question about the visitor’s own account, invoice or data is always declined, because Deacon can’t see it.

You can trace what it said. An answer taken from a page on your site usually links to that page, so your customer can read it. In your dashboard, every answer shows the source Deacon named, or “Your content” when it named none, and every conversation is saved word for word.

Grounding also shows you where your content runs out. Every question your content didn’t cover waits in Conversations with an Answer this button. Write the answer once and press Save & check, and Deacon shows you the reply it now gives and how many other open questions your answer covers. Every answer you write becomes part of what Deacon knows, so it needs you less over time. Topics group what customers ask and count the people behind each group, so you see what they’re stuck on and what they want next.

Every plan answers with the same model and follows the same rules, the free plan included, and the free plan needs no card and has no end date. How a knowledge base chatbot finds its answers shows each step on the dashboard’s own screens.

Questions about grounding AI

Is grounding the same as RAG?

Not quite. Grounding is the goal, answers tied to sources you chose. RAG is the most common way to get there, by retrieving passages and handing them to the model. A bot can use RAG and still be poorly grounded if nothing stops it using its own knowledge when the passages fall short.

Does grounding stop AI hallucinations?

It reduces them, but it doesn’t stop them. The legal research tools in the 2024 study above used retrieval and still made things up between 17% and 33% of the time. What makes a grounded support bot safe is what happens around the model, rules against filling gaps, a plain “I don’t know” and a record of every conversation you can read.

Does a grounded answer always cite its source?

Not always, and Deacon’s don’t. An answer from a page on your site usually links that page. An answer from a file, pasted text or a written answer shows the visitor no link, because there’s no public page behind it. Your dashboard still shows the source Deacon named.

Can I let Deacon answer from general knowledge?

No. There’s no setting for it, and your Instructions can’t switch it on, because Deacon is told they never override the rule to answer only from your content. A question outside your content gets an honest “I don’t know”, and one about something unrelated to your product gets a polite no. If Deacon keeps missing something customers ask, write the answer with Answer this.

What can still go wrong with Deacon’s answers?

Three things. If a page is out of date, Deacon repeats what it says. The search can pick the wrong passage, or the model can slip. And a written answer is found like any other source, so it can miss the cut for a question it should answer. When your answer and a page both reach the model and disagree, Deacon is told to prefer yours, and Save & check shows you whether Deacon now uses it.