AI hallucination — when the model confidently states something that's factually wrong or completely made up — is one of the most serious concerns with AI agents. The good news: hallucinations are rare with a well-configured ChatForge agent, and there are concrete steps you can take to reduce them to near zero.
What is hallucination and why does it happen?
Hallucination happens when an AI model generates a response that isn't grounded in any real information — it "fills in the gap" with plausible-sounding content that may be entirely false.
In a RAG-based system like ChatForge, hallucinations most commonly occur when:
- The knowledge base doesn't contain the answer, so the model tries to infer it
- The creativity (temperature) setting is too high
- The system prompt doesn't explicitly tell the agent to admit uncertainty
- There are contradictory entries in the knowledge base
Prevention 1 — Build a complete knowledge base
The most powerful hallucination prevention is giving the AI enough real content to answer from. When there's a clear, specific answer in the knowledge base, the AI uses it. When there isn't, it guesses.
Action: For every topic your agent should cover, make sure there's a specific entry in your knowledge base. Run your 20-question test and add content for every question that gets a "I don't know" answer — before the AI starts guessing.
Prevention 2 — Lower the creativity setting
The temperature setting controls how creative (and risky) the model is. Higher creativity = better fluency but higher hallucination risk. Lower creativity = tighter, more literal responses grounded in the source material.
- Recommended: 0.3–0.4 for fact-heavy business agents
- Avoid settings above 0.6 if accuracy is important
Prevention 3 — Explicitly instruct honesty in your prompt
This is one of the most underused and most effective tools. Add this to your system prompt:
Only answer questions based on information in your knowledge base. If you don't have specific information to answer a question, say: "I don't have that detail right now — please contact us at [email] or book a call: [link]" Never guess, infer, or make assumptions beyond what you know.
This single addition dramatically reduces hallucinations because the model is explicitly told not to speculate.
Prevention 4 — Remove conflicting content
Conflicting information in your knowledge base is a major hallucination trigger. If one entry says "delivery takes 3 days" and another says "5–7 days", the AI may confidently state either one — or even average them into a number that's completely wrong.
Action: Audit your knowledge base for any topic where you have multiple entries. Ensure they're consistent. Remove outdated entries immediately when your information changes.
Prevention 5 — Set a strong fallback message
The fallback message is your safety net. When the agent can't find a confident answer, it should always route to a human contact instead of guessing. Configure your fallback to include your email, phone, or booking link so the visitor still has a path forward.