#AIGC技术内容创作征文｜全网寻找AI创作者，快来释放你的创作潜能吧！#
Article Directory
Exclusive question answerer for personal blog: demo
1 Introduction
GPT is an acronym for Generative Pre-Training Transformer. It is a family of large-scale language models developed by OpenAI. The models of the GPT series are based on the Transformer architecture and use the attention mechanism to selectively focus on relevant parts of the input text. GPT models are trained using large amounts of data to predict the next token in a sequence and are able to generate coherent and contextually relevant text.
The most recent model in the GPT family is GPT-4, released on March 14, 2023. GPT-4 is a multimodal language model, meaning it can accept both text and images as input. It has been fine-tuned with reinforcement learning and AI feedback to ensure consistency with humans and follow policy.
GPT-3 is another notable model in the series, released in 2020. It had 175 billion parameters, making it one of the largest language models available at the time. GPT-3 demonstrates strong zero-shot and few-shot learning capabilities on various tasks.
OpenAI has also developed a chatbot called ChatGPT, which is built on the GPT model. ChatGPT allows users to have interactive conversations and provides options to adjust and guide the conversation to the desired length, format, style, level of detail and language used. It has become very popular and offers a free version and a paid subscription service "ChatGPT Plus".
As we all know, the knowledge learned by GPT3.5 will end in 2021, and the answers to things after that are all made up. We design two application scenarios. One is that if you want to use GPT to create a software platform dedicated customer service, if you only access ordinary GPT3.5, it is basically impossible to answer the questions raised by users in a targeted manner. The second is that when we are doing academic work, we often read some cutting-edge articles, which are not learned by GPT. Then we ask GPT about the content of this article, and GPT will undoubtedly make it up.
So, we will solve this problem today, how to make the answer of GPT3.5 more targeted and accurate in specific fields. I will explain this through two experiments:
(1) Build my blog question answering system , the GPT in this system can explain my blog content to users and answer users' questions about my blog. The link of the system is as follows: demo ;
(2) Use targeted GPT to help us read cutting-edge papers;
2 Build my blog question answering system
2.1 Introduction to the platform (accessible without magic)
Dify.ai : The word Dify comes from Define + Modify, which means to define and continuously improve your AI application, which is done for you (Do it for you).
If you're amazed and excited by the rapid development of LLM techniques like GPT-4, you can't wait to do something useful with it! But you have all the puzzling questions in your head:
- How do I "train" a model based on my content?
- How to let AI know what will happen after 2021?
- How to avoid letting AI talk nonsense to users?
- What do Fine-tuning and Embedding mean?
Well, Dify is just what you need.
The goal of Dify is to allow developers (even non-developers) to quickly build useful things based on large language models, and ensure that it is visualized, operable, and improveable.
2.2 Early preparations
Since the exclusive GPT created is to learn your blog post knowledge, you need to export your blog post here, and the considerate CSDN provides us with an export function!
Click Export on the blog editing page → click Export as a MarkDown file
ps: This can also be used as a backup method for your blog
After exporting all your blogs, organize them into a folder, and the preliminary preparation is over!
2.3 Upload the blog as a dataset to Dify
Step1: First you need to log in to Dify, click on the dataset on the home page:
step2: Click on the dataset page to create a dataset:
step3: Click on the interface to create a dataset to create an empty dataset:
step4: Next, create a dataset for your Give the blog dataset a name (I named it blog here):
Step5: Import all the blog files of your question md into the blog dataset folder, and click Next:
Step6: Click "Save" as shown in the figure below and process" (in fact, the high-quality indexing method is more recommended here):
step7: Wait for the system to complete the processing, that is, all files are displayed as available:
So far, your blog post knowledge has been passed into the system as a data set!
2.4 Build your blog answer specialist
Step1: We come to the build application area and click Create Application:
Step2: We create a name for this application and select a conversational application:
step3: browse the application information, you can check the application URL, user active number and other information on the page you just entered, which is convenient for future management: step4:
click the prompt word arrangement to start the robot setting, first you need to "hypnotize" GPT, let him become Your blog commentator, set the following settings in the "prompt words before dialogue" column; then you need to upload your blog knowledge, and select the data set you just uploaded in the context: step5: After checking everything is correct, click Publish, and
then Visit the URL that appears in step3 to use your exclusive blog commentator!
The core here is to allow GPT to bring in the identity before the user talks to GPT, and "teach" GPT specific knowledge, so that GPT will not make up random questions about your blog raised by users, so as to carry out perfect answer!
ps: In the same way, this method can also be used to build the exclusive customer service of the software platform, and the help document of the platform needs to be used as the knowledge base~
3 Use targeted GPT to help us read cutting-edge papers
The core principles and operations here are mainly the same as those in Chapter 2. First, you need to upload the PDF version of the paper as a dataset to Dify:
We are on the homepage of Dify, click "Smart Chat", and the following interface will appear:
You can see that Dify has integrated the function of GPT networking, and we will input the paper just uploaded as the context to GPT:
Now you can use GPT to ask specific papers content, and the platform will automatically generate possible next questions based on your conversation with GPT, which makes it very convenient to obtain information.
As you can see from the picture above, this paper was published in 2023. Let's try it out, what kind of answer will GPT give if we don't give the prerequisites of the GPT context?
GPT is a large-scale language model trained based on the GPT-3.5 model, but it should be noted that its knowledge is only as of 2021. Therefore, for questions involving the latest information or related to the original text of the paper, GPT will only be able to make up nonsense and cannot provide accurate and professional answers.
4 Epilogue
The construction of blog question answering system is very useful. Using GPT, the system can explain blog content and answer users' questions. Through the Dify.ai platform, we can easily define and continuously improve our own AI applications. This allows us to take advantage of the latest LLM techniques, including models like GPT-4. The platform also clears up some of the confusion about training models, understanding what's beyond 2021, and avoiding nonsense. It provides fine-tuning and embedding capabilities that allow us to better manage and optimize our applications.
In short, Dify brings us the convenience and flexibility of building a blog question answering system. By utilizing the GPT-3.5 model and the Dify.ai platform, we can provide users with more in-depth and accurate answers, thereby improving user experience and ensuring the correct delivery of information. It's a compelling and useful app that will be a valuable and powerful tool for a personal blogger or other content provider.
If you like it, please like and comment
