The problem when working with llms. They are a problem for two main reasons. The first reason is the hallucination naturally causes the user to receive an incorrect answer. Second, the opposite bad reason, halucinations reduce the trust of users in the system. Without a user who believes in your answer to the question, it will be difficult to keep users on your platform.
Content
Why do you need to reduce Hallucinations?
When clients think through the llm to solve the problem, one of the first thoughts that often get to mind is the Lytonanitions. It used to hear that the LLMs sometimes issue a fake text, or answers at least you cannot trust completely.
Unfortunately, they often direct, and you need to take steps to reduce HALLucinations in your answer to the Systems. The contents of this article will refer to the Hallucinations directly to access the question response system, although all strategies can be used for other LLM programs and, such as:
- To schedule a particular type
- Release of information
- Default
Throughout this article, I will discuss real strategies in the world You can apply to reduce Halkinations' impact, can prevent them from slow or reduce harmful damage. This, for example, can hurt as a user to trust your program slowly after dealing with it
HALLucinations Prevention Strategies
I will distinguish this section of two clauses:
- Strategies Direct Limit The number of Hallucinations you receive in llms
- Strategies to reduce Hallucinations damage
I did this because I think it helps to get the techniques that you can use to reduce the impact of halucinations. You can try to prevent them from happening at all.
Reduce the number of halucinations
In this clause, I will cover the following strategies:
- A verification step (llm judge)
- RAG development
- Prompt Immediate Your Program
Llm is verification of verification
The first way I will cover uses the llm as a judge to verify your llm responses. This method depends on the mind below:
To ensure feedback, usually a simple employee than producing an answer
This quotation is sometimes easily understood with mathematical problems, where the solution is usually difficult, but ensures that the solution is correct. However, this concept also works in your answer to the question.
To remove the answer, the llm should read a lot of text and translate the user's question. The llM then should come with the appropriate response, submitted to the context of the FEDD. To ensure an answer, however, it is often easier, consideration of the llm confirmation requires judgment that the final answer is given the question and context. You can learn more about the llM verification in my article with a great vindication of the LLM production.
RAG development
There is also a lot of progress you can do in your RAG pipe to protect the halucinations. The first step is to track the correct documents. I just wrote about this process, with the strategies to raise both the increasing accuracy and remembering documents that downloaded the RAG. Starts it boots down to filter out the wrong documents (increasing specification) with a reconanking and llm verification. On the other hand, you can confirm that you entered the relevant documents (up to remember), with strategies such as the restoration of the content and download multiple chunks.
Prompt Immediate Your Program
Another process you can use to reduce the number of halucinations to quickly improve your system. Anthropic has just wrote an article about writing AII agents, highlighting how they use Claude code to do its issuance. I recommend to do something similar, when feeding all your billing, asking the llm to improve your world general, and highlight charges of success.
In addition, you must enter a sentence about your prominence soon that the llm should only use the information given to answer the user's question. He wants to block the model to reach the birthday based on previous training, and instead use the destination given.
prompt = f"""
You should only respond to the user question, given information provided
in the documents below.
Documents: {documents}
User question: {question}
"""This will significantly reduce the effect that the llm responds based on previous training, which is the normal source of vblams.
Reduce Halkinations
In a previous clause, I covered the plans to prevent around it. However, sometimes the llm is still about to postpone, and you need ways to reduce damage when this happens. In this clause, I will cover the following strategies, helping to reduce the impact of the Halkinations:
- To close your sources
- Help the user well using the system
To close your sources
One powerful plan to make a llm deliver its sources when it gives answers. To do, for example, see this whenever you ask Chatgt to answer the question, based on the Internet content. ChatGPT will provide you with the answer, and after the answer text, you can see the display of citation citation on the website where the answer was taken.
You can do the same with your RAG program, either live while answering a question or working in the background. To live, for example, you can, for example, provide each IDs in each document document, and ask the llm to issue any document. If you want high quality quotes, you can do in the background, which you can learn more about the Anthropic Docs.
To direct the user
I also think is appropriate to say how to guide the user to make good use of your app. You, as the Creator of your question to answer a question, well knowing their strength and weaknesses. Perhaps your plan is amazing in answering one type of question, but it makes the worst of some forms of questions.
When this is the case (most commonly), I am very commending you to inform the user of this. The opposition argument here is that you do not want the user to know the weakness of your system. However, I can argue that it is better to appreciate the user of this weakness, rather than the user deal with weaknesses by themselves, for example, through the connection.
Therefore, you must have a document that surrounds your order to answer a question, or onbearding, which is notifying the user that:
- The model is very effective, but occasionally can make mistakes
- What kind of model questions are ready to answer, and what kind of questions do you work in
Summary
In this article, I have negotiated HALLucinations for llms. Halloucinations is a serious problem, which is many that many users know. You should always have several direct steps at the Halkucinations reductions. I have said that obvious strategies reduces halublions, such as improving the RAG document in the account, performing your System program, and the llM verification. In addition, I also discussed how you can reduce the damage to HALLucinations if they happen, for example, by giving users to sources and guide the user to effect your application.
👉 I have found in the community:
🧑💻 Contact your
🔗 LickDin
🐦 x / Twitter
✍️ Medium
