Stop the chatbot from making things up.
Wiki cannot guarantee no hallucinations. It requires source-backed answers, refuses when nothing is approved, and keeps unverified or internal content out of public bots.
Wiki cannot guarantee no hallucinations. It requires source-backed answers, refuses when nothing is approved, and keeps unverified and internal content out of public bots.
Chatbots make up answers when nothing approved is in reach.
A chatbot making up answers is rarely a model surprise. It is what happens when retrieval returns a near miss, an old file, or an internal note, and the prompt still says to be helpful.
Teams try to stop chatbot hallucinations with longer prompts, more evals, or a bigger index. Those help you observe the failure. They do not decide which documents are reviewed, which collections a public bot may use, or when the product should refuse.
The honest control is narrower. Require a verified source. Keep internal collections off the public application. Test the questions that already caused damage. Refuse when the approved set has nothing to cite.
What to do instead of hoping the model behaves
Treat hallucination as a knowledge and policy problem first. Prompt work comes after the source is governed.
- Step 1
Decide what the public bot is allowed to know
List the collections that are public-safe. Everything else stays out of that application's policy, including helpful internal write-ups.
- Step 2
Verify the sources that application may cite
If nobody reviewed the page, the bot should not speak from it. Verification is the gate, not a label you add later.
- Step 3
Require a citation or a refusal
A source-backed answer is the success case. No approved source is a refusal, not a best guess.
- Step 4
Test the questions that already went wrong
Replay them in the Answer Playground. If the bot still invents a number, the collection or the policy is wrong.
How Wiki reduces made-up answers, honestly
This is not a guarantee. It is the live control surface for sources, collections, policies, and proof.
Verified sources only
Unverified documents do not become retrieval material for an application that requires review.
Public-safe versus internal split
Internal collections stay off public chatbots. That removes a common source of confident, wrong, internal leakage.
Permission-aware retrieval
The AI Application policy is the allow-list. The model does not get to rummage the whole workspace.
Answer Playground
See the citations, or the refusal, before a customer does.
Wiki Connect
Your bot calls an HTTP API that already applied the policy. You are not re-implementing the allow-list in the prompt.
Questions teams ask first.
- Can Wiki guarantee no chatbot hallucinations?
- No. Wiki cannot guarantee no hallucinations. It requires source-backed answers, refuses when nothing is approved, and keeps unverified or internal content out of public bots.
- Why does my chatbot keep making up answers?
- Usually because it can retrieve draft, stale, or internal documents, or because the prompt still asks it to answer when no approved source matches. Close the collection, verify the source, and test a refusal.
- Is this just RAG with a nicer UI?
- No. RAG retrieves. Wiki decides whether a document is verified, which application may use it, and whether the answer must cite or refuse. Retrieval without that layer is not a source of truth.
- How do I try this on a live refund question?
- Follow the refund policy walkthrough. Import the public URL, verify it, put it in a public-safe collection, attach a policy, and test the question in the Answer Playground.
Essays, tools, and the product.
From the blog
Why Your AI Chatbot Keeps Giving Different Answers
Inconsistent chatbot answers usually come from fragmented sources, stale docs, conflicting policies, and missing governance, not random model behavior.
RAG Is Not a Source of Truth
Retrieval can find context. It does not decide whether that context is current, approved, allowed for the audience, or safe to cite.
The Complete Guide to Citation-Backed AI Support
Customer-facing AI should show where the answer came from, know when it lacks an approved source, and escalate when documentation cannot support a reliable answer.
Tools
Your AI is already talking.
Make sure it knows what it is talking about.