After the launch of ChatGPT-related APIs, various tools have sprung up like mushrooms after the spring rain. This article lists some tools that are used daily.
tool list
myreader.io
myReader Homepage
The author of this tool is @madawei2699, the github home page address is t.co/adJBYWbjkF,…
- Read any web content online, including videos (YouTube), and answer your relevant questions or summarize relevant content based on these content
- Support reading e-books and documents (support PDF, EPUB, DOCX, Markdown, TXT), and answer your relevant questions or summarize relevant content based on these contents
- Regularly send daily hot list news, no matter the news is in Chinese or other languages, it can use chatGPT to automatically summarize the content of the news in Chinese, so that you can quickly obtain hot news information
- Supports prompt templates, can answer your questions based on the context of the message history, and even play games with you
- Support multi-language voice interaction (English, Chinese, German and Japanese), it will answer your questions with the voice of the relevant language according to your language, so as to help you train your foreign language ability, it can be understood as your private foreign teacher
For specific function demonstrations, please refer to my AI reading assistant .
chatpdf
ChatPDF Home
This can be regarded as a tool for assisting PDF reading. After users upload their own PDFs, they can interact with the tool in a dialogue manner to quickly obtain the content of the PDF file.
Chat Paper
A tool that focuses on "scientific research dogs", summarizes papers through ChatGPT, and helps "researchers" conduct preliminary screening of papers (currently does not support dialogue on the content of papers).
ChatPaper Homepage
In addition, related tools include polishing tools, review tools, and review reply tools .
Finally, the classification of chatGPT-based application types mentioned in this article- ChatGPT Application Development Notes is also of reference.
principle
I was planning to write an introduction to the principles of the above-mentioned tools (in fact, ChatGPT’s plug-in - chatgpt-retrieval-plugin), but after checking the source code of several projects, I found that the main principles of such tools are actually quite intuitive:
- Parse relative input as text
- Obtain the embedding of the sentence after dividing the text into sentences (the current processing method of this step is mostly truncated according to the length) and store it in the database
- User input is converted to embedding, and the most relevant sentence set is recalled in the database
- Assemble the recalled sentence and the user input sentence as the input of ChaptGPT, and obtain the output
Although the above ideas are intuitive, in order to obtain better results, in fact, except for the third step, each step has room for optimization:
- Text analysis can be used for targeted analysis of different types of data
- The text segmentation method can adopt special punctuation for sentence segmentation, and there are many optional generation methods for sentence embedding
- The recalled sentence and the user input sentence are assembled as the input of ChaptGPT, combined with the task-specific prompt, to obtain an output that is more suitable for the task
The specific flow chart can refer to gpt-langchain-pdf :
gpt-langchain-pdf