检测与缓解幻觉
幻觉是 LLM 自信生成却不真实的输出。学习幻觉产生的原因、检测方法，以及在生产环境中减少幻觉的具体技术。
检测与缓解幻觉 是 CoddyKit 上的免费 Prompt Engineering & LLM Optimization for Developers 课时。 这是第 4 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Prompt Engineering & LLM Optimization for Developers 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Prompt Engineering & LLM Optimization for Developers 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
What is a Hallucination?
A hallucination is when an LLM produces text that is fluent and confident but factually wrong or unsupported by any source.
Because the output looks authoritative, hallucinations are dangerous in user-facing apps.
Why They Happen
LLMs predict the most likely next token, not the truth. With no grounding, they fill gaps with plausible-sounding inventions.
- Missing knowledge in training data
- Ambiguous or leading prompts
- Pressure to always answer
Types of Hallucination
Two broad categories:
- Factual: wrong dates, fake citations, invented APIs
- Faithfulness: answer contradicts the provided context (common in RAG)
Grounding with Context
The strongest defense is grounding: give the model trusted source text and instruct it to answer only from that text.
Answer ONLY using the context below.
If the answer is not in the context, say "I do not know".
Context:
{retrieved_docs}Forcing Citations
Ask the model to cite which passage supports each claim. Unsupported sentences become easy to spot and verify.
For each sentence, add a [doc_id] citation.
Do not make claims you cannot cite.Lowering Temperature
Higher temperature increases creativity — and invention. For factual tasks, set a low temperature so the model stays close to high-probability, well-grounded tokens.
const res = await client.chat.completions.create({
model: "gpt-4o-mini",
temperature: 0,
messages
});Self-Consistency Checks
Generate the answer several times. If the model gives different facts each run, the claim is likely a hallucination. Agreement is a weak but useful signal of reliability.
LLM-as-a-Judge Verification
Use a second model call to check whether the answer is supported by the context. The judge returns a faithfulness verdict you can act on.
Does the ANSWER follow only from the CONTEXT?
Reply: SUPPORTED, PARTIAL, or UNSUPPORTED.Programmatic Validation
When outputs are structured, validate them. A cited URL should resolve, a quoted number should match the source, a JSON field should match a schema.
function validateCitation(cit, docs) {
return docs.some(d => d.id === cit.doc_id);
}Letting the Model Abstain
Give the model permission to say I do not know. Removing the pressure to always answer measurably reduces fabricated content.
Human Review for High Stakes
For medical, legal, or financial outputs, route low-confidence or uncited answers to a human before showing them to users. Automation plus oversight beats either alone.
Quick Check
Test your understanding.
Recap
You learned to fight hallucinations: ground answers in trusted context, force citations, lower temperature, use self-consistency and LLM-as-a-judge checks, validate outputs programmatically, allow abstention, and add human review for high-stakes cases.
常见问题解答
「检测与缓解幻觉」课时是免费的吗？
是的 — 「检测与缓解幻觉」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Prompt Engineering & LLM Optimization for Developers 课程的其余内容，请升级到 CoddyKit PRO。 Prompt Engineering & LLM Optimization for Developers 课程共包含 4 节课。
「检测与缓解幻觉」这节课中我会学到什么？
幻觉是 LLM 自信生成却不真实的输出。学习幻觉产生的原因、检测方法，以及在生产环境中减少幻觉的具体技术。 你通过在浏览器中直接运行的动手代码来练习 Prompt Engineering & LLM Optimization for Developers，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 Prompt Engineering & LLM Optimization for Developers 需要有经验吗？
无需任何先前经验。CoddyKit 上的 Prompt Engineering & LLM Optimization for Developers 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 4 节课，共 4 节。
「检测与缓解幻觉」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 Prompt Engineering & LLM Optimization for Developers 课中编写并运行代码吗？
能。每节 Prompt Engineering & LLM Optimization for Developers 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。