/articlesComparisons
Answers from your sources vs a general model
A general AI model guesses your prices and policies. A grounded assistant answers from your files and systems. How it works, where it fails and how to test it.
- Updated
- Published
- min read
- 05
- On this page
- 08
On this page08
- What does “grounded” mean for an assistant?
- Why does a general model make things up about your company?
- Grounded assistant vs general model, side by side
- How do you ground an assistant?
- Does grounding stop all wrong answers?
- How to test whether an assistant is grounded
- How Botify grounds its assistants
- Common questions
01
What does “grounded” mean for an assistant?
Grounded means every answer comes from something the assistant looked up for this question: a line in your price file, a page on your website or a record in your system.
A grounded assistant has three habits. It looks things up before it answers. It answers only from what it found. And when nothing it found covers the question, it says so or passes the conversation to your team. Most demos skip the third habit, and most real mistakes come from it.
02
Why does a general model make things up about your company?
A general model has never seen your refund policy, your price list or today’s stock. Ask it anyway and it writes the most likely-sounding answer. Likely-sounding and correct are different things, and the model cannot tell which one it wrote.
This has cost companies money. In February 2024 a Canadian tribunal held Air Canada responsible after its website chatbot told a customer he could claim a bereavement fare after travelling, which the airline’s policy did not allow. The tribunal said the airline is responsible for everything on its website, whether it comes from a page or a chatbot.
Whatever answers for your company should answer from your company’s sources. A polite wrong answer is still your answer.
03
Grounded assistant vs general model, side by side
A grounded assistant still uses a general model inside. The model understands the question and writes the reply; your sources decide what is true.
| General model | Grounded assistant | |
|---|---|---|
| Where the answer comes from | What it learned in training | Your files, website and systems, looked up at question time |
| Your prices and policies | Guessed, often convincingly | Read from your price file or system, or not answered |
| How current | Fixed at its training date | As current as the file or system it reads |
| When there is no answer | Writes one anyway | Says it does not know or hands over to a person |
| Checking an answer later | No source to check | Each answer traced to a file or a system lookup |
| Changing a fact | Wait for a new model | Edit the file or the record |
| Best for | Drafting, brainstorming, translation | Customer answers, quotations, bookings, anything you are liable for |
04
How do you ground an assistant?
There are three ways, and most real setups use all three. The mix depends on where your answers live.
- Files and web pages: the assistant searches your documents, such as a price list or a clinic’s services page, and answers from the passages it finds. This is often called RAG.
- Live systems: the assistant asks the system directly, such as Odoo for today’s price and stock, so the answer matches what your team sees.
- Exact fields: for numbers such as an order total or an appointment time, the assistant reads the field instead of estimating it.
- Instructions around all three: answer only from what you found, and when nothing fits, say so or hand over to a person.
05
Does grounding stop all wrong answers?
No. It changes the question from “the model does not know” to “did the assistant find and use the right source?”. That is much easier to see and fix, but it can still go wrong.
- A missed search: the right page exists, but the assistant found a weaker one.
- Old or conflicting files: last year’s price list is still uploaded next to this year’s.
- Mixing in guesses: the model adds its own assumptions to what it found.
- Planted instructions: a document or message contains text aimed at the model; the assistant should treat it as data, never as an order.
- Answering next to the topic: the assistant answers a question its sources do not quite cover instead of passing it on.
06
How to test whether an assistant is grounded
Write 20 to 30 test questions before launch and run them again after every change to instructions, files or models.
- Questions your files answer: the reply should be right and match the source.
- Questions your files do not answer: the assistant should say so, not improvise.
- Questions where two files disagree: it should flag the conflict or use the file you marked as current.
- A price you just changed: the next answer should use the new price.
- The same question in Arabic, in Gulf dialect and in English: the answers should agree.
07
How Botify grounds its assistants
In Botify you upload PDF, Word, Excel, CSV, HTML, Markdown and text files, or point the assistant at your website; it reads JavaScript-heavy sites too. It searches them by keywords and by meaning together, so a customer who writes “كم سعر التنظيف؟” still finds your teeth-cleaning price.
For live data the assistant works in your systems, such as checking a product’s price and stock in Odoo before it prepares a quotation. Every lookup is written to the audit log, so you can see what an answer was built on.
When a question is outside what it knows, the assistant can hand the conversation to a person on your team. Before you publish a new version, Botify can run your saved test questions.
08
Common questions
01Can a grounded assistant still answer general questions?
You can allow it, but it is usually better to steer back to your business. If you allow general answers, make sure customers can tell a company fact from general knowledge.
02Is grounding the same as RAG?
RAG, retrieval-augmented generation, is one way to ground an assistant: it searches your documents and answers from the passages it finds. Grounding also covers asking live systems and the instruction to stay silent or hand over when nothing fits.
03Does training a model on our data make it grounded?
Not by itself. Fine-tuning can change a model’s tone and vocabulary, but facts learned that way come from memory, cannot be traced to a source and go stale. Keep facts in files and systems the assistant looks up.
04How do I update what a grounded assistant knows?
Update the source: replace the price file, edit the page or fix the record in your system. The next question uses the new version, with no retraining. Keep one current version of each file so the assistant never finds two.
09
Sources
10