Why this matters
A support bot that guesses is worse than no bot. Ask a plain chatbot what your firm charges and it will give a typical price, confidently, because that is what it learned in training. One wrong fee quoted on your website costs more than every question the bot answers correctly.
Every bot in this course fixes that the same way: it answers from your own documents. Chatbase calls them Data sources, a Custom GPT calls them Knowledge, and in n8n you load them into Pinecone yourself. The idea underneath is called retrieval, and once you understand it, you know why a bot gets an answer right, and what to fix when it does not.
What you’ll build
You follow one customer question through a website bot for Alder & Finch Accounting, a made-up five-person firm: three files cut into chunks and indexed, a search by meaning with match scores, the AI told to answer only from what it found, and a question the files do not cover that gets a fallback answer instead of a guess.
What you need
- Nothing but a browserIt is a short explainer of diagrams you click to open. No account, no files and nothing to set up.
Concepts
3 ideas you’ll use in the practice.
Upload once: chunks and an index
When you add a file, the tool cuts it into chunks of a few sentences and turns each chunk into an embedding, a list of numbers that captures its meaning. The chunks are kept in a search index, often called a vector database. Change a price in the file and you upload it again, or the bot keeps the old one.
Every question: retrieve, then answer
The question is matched to the closest chunks by meaning, each with a score. Only the best few go to the AI, together with a rule to answer only from them. That is retrieval-augmented generation, or RAG. A good bot also names its sources, so you can check them.
Weak matches mean "I don’t know"
When nothing matches well, the bot should give a fallback answer that points to a person. Testing one question your files do not cover is the quickest way to see whether a bot guesses.
Practice
Do every step yourself in a practice copy of the real tool. It checks your work, and “Do it for me” finishes any step you get stuck on.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Quick check
3 questions to check you’ve got it.
1. Why does a bot with your documents quote your real fee instead of a typical one?
- It was retrained on your website
- It finds the matching chunk of your fee sheet and answers only from it
- It searches Google for your firm
Show the answer
It finds the matching chunk of your fee sheet and answers only from it. Retrieval finds the closest chunks of your own files and the AI is told to answer only from them.
2. A visitor asks something none of your files cover. What should happen?
- The bot answers from general knowledge
- The bot gives a fallback answer that points to a person
- The bot picks the closest chunk anyway
Show the answer
The bot gives a fallback answer that points to a person. Weak matches are the signal to stop. A fallback answer is safer than a confident guess.
3. Your bot quotes last year’s price. What do you fix first?
- The file it learned from
- The bot’s name
- The colour of the chat bubble
Show the answer
The file it learned from. Retrieval reads your files on every question, so the fix is the file: update it and upload it again.
Chatbase: Answer Customer FAQs with an AI Agent
You know how answering from your docs works. Next, build it for real: a Chatbase agent trained on a firm’s FAQ, with a fallback answer.
Start module 4 →