If your agent is guessing, silent or missing from your site, start with the table.
What are the most common Chatbase FAQ agent mistakes?
| # | Mistake | Fix |
|---|---|---|
| 1 | No fallback rule in Guardrails | Add a rule: if the answer is not in the training data, say so and give your email and phone |
| 2 | Uploading a scanned PDF | Use a PDF with selectable text, or a .txt, .doc or .docx file |
| 3 | Training on the website when the answers live in a file | Leave the website box empty and click Add Notion, Files, Text and more |
| 4 | Keeping the default General text | Replace it with your business, the agent's role and its tone |
| 5 | Pushing Temperature toward Creative | Leave it at Reserved for prices and policies |
| 6 | Forgetting Save changes | Nothing you type in Instructions is used until you save |
| 7 | Going live before testing in the Playground | Ask real questions plus one the FAQ does not cover |
| 8 | Bubble left on Channel disabled, or Specific domains off | Click Enable chat bubble, then lock the embed to your own site |
| 9 | Never reading the chat logs, or letting the FAQ go stale | Read Activity weekly, use Revise answer, update the source file |
Why does my Chatbase agent guess prices or policies?
Mistakes 1, 4 and 5 all point the same way: the agent has no clear rule for what it may say.
1. No fallback rule. Chatbase starts you with two Guardrails, No Data Disclosure and Maintaining Focus. Neither tells the agent what to say when the FAQ has no answer. The lesson adds three rules, and the one that matters most is the fallback answer: "If the answer is not in the training data, say: I'm not sure about that one. Please email hello@juniperlane.example or call (555) 014-2290 and a person will help." That single rule is the agent's human handoff. The rule just before it says to answer only from the training data and never guess prices, dates or policies.
4. Leaving the default General text. The General box opens with a generic role plus language and formatting rules. The agent does not know it works for a bookkeeping firm that does not prepare tax returns until you say so. The lesson's version names the business, lists what it does and does not do, and says "Use only our FAQ. If a question is not clear, ask one clarifying question."
5. Raising the temperature. Model configuration has a Temperature slider from Reserved to Creative, and it starts at Reserved. Reserved keeps answers focused and consistent. For a price list or a refund policy, that is exactly what you want, so leave it.
Why didn't my FAQ file train the agent?
2. A scanned PDF. The Add File panel accepts pdf, doc, docx and txt, and shows a yellow note that PDF text must be selectable. A scan of a printed price sheet is a picture, not text. Export it as a real PDF or paste it into a .txt file. The lesson's sample FAQ is a plain .txt file of 4 KB.
3. Training on the website by default. The first set-up screen asks How would you like to train your AI Agent? and suggests your website. If your answers are in a document, leave that box empty, click Add Notion, Files, Text and more, then File. Afterwards, open Build > Data sources and confirm the file is listed with Last synced at the top. If it is not there, the agent has nothing of yours to answer from.
Why aren't my instruction changes showing up?
6. Skipping Save changes. It sounds obvious, but it is easy to edit General and Guardrails, jump to the Playground and wonder why nothing changed. Click Save changes at the top of Instructions. Chatbase confirms with Your changes are saved.
How do I know the agent is ready to go live?
7. Skipping the Playground. The Playground shows the agent the way a visitor sees it. Ask the questions customers really send and check each answer against the FAQ. In our run, "Are you open on a Saturday in March?" got "Yes. In March, we're open Saturdays from 9:00 AM to 1:00 PM Eastern," which matches the file. Then ask one question the FAQ does not cover. "Do you work with nonprofits?" got the fallback, word for word. If yours invents an answer instead, tighten the Guardrails and test again before you touch the embed code.
Try it free, step by step
Chatbase: Answer Customer FAQs with an AI Agent. A business case for a small firm that answers the same customer questions every week: build a Chatbase AI agent on your FAQ in the real app, make it say when it does not know, test it, get the chat… You do every step yourself in a practice copy of Chatbase, it checks your work as you go, and “Do it for me” finishes any step you get stuck on. Free, and nothing touches your real accounts.
Start the free lessonWhy isn't the chat bubble on my website?
8. The bubble is off, or the embed is open to anyone. After Deploy > Website widget, the Website bubble panel says Channel disabled. The bubble stays off until you click Enable chat bubble. Then turn on Specific domains ("Only allow embedding the agent on specific domains") and add your own domain, one per line. Without it, anyone could paste your agent onto their site. Only then copy the script under Widget setup, the embed code your web person pastes in. Send instructions via email sends it to them for you.
What should I do after the agent is live?
9. Ignoring the logs and the source file. Activity > Conversations holds the chat logs, and every answer has a confidence score. In our run the Saturday answer scored 0.755 and the nonprofit fallback 0.369, shown in red. Start with the low scores. Click Revise answer, add a Q&A Title, type the right reply in Expected response and click Update Answer. Chatbase saves it as a new Q&A source, and a fresh chat in the Playground used it straight away.
The other half of this mistake is the FAQ itself. When a price or policy changes, update the source. The agent says what its knowledge base says, so an old file means old answers. The lesson suggests 15 minutes a week: read the chats, fix two or three answers, add what is missing.
Key takeaways
- Add a fallback rule to Guardrails before anything else. It is the difference between "I'm not sure" and a made-up price.
- Use a file Chatbase can read, and confirm Last synced on Data sources.
- Replace the default General text, keep Temperature at Reserved and click Save changes.
- Test real questions and one unknown question in the Playground before you copy any code.
- Enable the chat bubble and turn on Specific domains.
- Read the chat logs weekly and keep the FAQ file current.
For the full set of questions about how it works, see Chatbase Questions Answered. The one-page cheat sheet has every setting on one printable page, and Chatbase's entry in the AI tools directory sums up what it does.
Frequently asked questions
- Why does my Chatbase agent make things up?
Usually because it has no rule against it. Add Guardrails that say to answer only from the training data and to give a fallback answer for anything else, keep Temperature at Reserved, then retest a question your FAQ does not cover.
- Can I fix one wrong answer without editing the whole FAQ?
Yes. Open the conversation in Activity, click Revise answer under the answer, type the right reply in Expected response and click Update Answer. It becomes a new Q&A entry in your data sources.
- Which answers should I check first in the chat logs?
The ones with low confidence scores. Chatbase shows a score on each answer and highlights low ones in red. In our run, the fallback answer to an unknown question was the low one.
- Do I need the tech stack screen to get started?
No. In the lesson, Select the tools in your tech stack is only a note for later. Pick nothing and click Continue.
- Is it safe to ask customers for documents in the chat?
The lesson adds a Guardrail so the agent never asks for bank statements, passwords or Social Security numbers in the chat. The sample FAQ also tells clients to use a secure portal instead of email.
